The Computational Theory of Mind (2015)

1. Turing machines

The intuitive notions of computation and algorithm are cardinal to mathematics. Roughly speaking, an algorithm is an explicit, step-by-step process for answering immoderate mobility or solving immoderate problem. An algorithm provides routine mechanical instructions dictating really to proceed astatine each step. Obeying the instructions requires nary typical ingenuity aliases creativity. For example, the acquainted grade-school algorithms picture really to compute addition, multiplication, and division. Until the early twentieth century, mathematicians relied upon informal notions of computation and algorithm without attempting thing for illustration a general analysis. Developments successful the foundations of mathematics yet impelled logicians to prosecute a much systematic treatment. Alan Turing’s landmark insubstantial “On Computable Numbers, With an Application to the Entscheidungsproblem” (Turing 1936) offered the analysis that has proved astir influential.

A Turing machine is an absurd exemplary of an idealized computing instrumentality pinch unlimited clip and retention abstraction astatine its disposal. The instrumentality manipulates symbols, overmuch arsenic a human computing supplier manipulates pencil marks connected insubstantial during arithmetical computation. Turing says very small astir the quality of symbols. He assumes that primitive symbols are drawn from a finite alphabet. He also assumes that symbols tin beryllium inscribed aliases erased astatine “memory locations”. Turing’s exemplary useful arsenic follows:

• There are infinitely galore representation locations, arrayed successful a linear structure. Metaphorically, these representation locations are “cells” connected an infinitely agelong “paper tape”. More literally, the representation locations mightiness beryllium physically realized in various media (e.g., silicon chips).
• There is simply a cardinal processor, which tin entree 1 representation location at a time. Metaphorically, the cardinal processor is a “scanner” that moves on the insubstantial portion one “cell” astatine a time.
• The cardinal processor tin participate into finitely galore machine states.
• The cardinal processor tin execute 4 simple operations: write a awesome astatine a representation location; erase a awesome from a memory location; entree the adjacent representation location successful the linear array (“move to the correct connected the tape”); entree the previous memory location successful the linear array (“move to the near connected the tape”).
• Which simple cognition the cardinal processor performs depends entirely upon 2 facts: which awesome is presently inscribed astatine the present representation location; and the scanner’s ain existent machine state.
• A machine table dictates which simple cognition the central processor performs, fixed its existent instrumentality authorities and the symbol it is presently accessing. The instrumentality array besides dictates how the cardinal processor’s instrumentality authorities changes fixed those same factors. Thus, the instrumentality array enshrines a finite group of routine mechanical instructions governing computation.

Turing translates this informal explanation into a rigorous mathematical model. For much details, spot the introduction on Turing machines.

Turing motivates his attack by reflecting connected idealized human computing agents. Citing finitary limits connected our perceptual and cognitive apparatus, he argues that immoderate symbolic algorithm executed by a quality tin beryllium replicated by a suitable Turing machine. He concludes that the Turing instrumentality formalism, contempt its utmost simplicity, is powerful capable to seizure each humanly executable mechanical procedures complete symbolic configurations. Subsequent discussants have almost universally agreed.

Turing computation is often described arsenic digital alternatively than analog. What this intends is not ever truthful clear, but the basic idea is usually that computation operates complete discrete configurations. By comparison, galore historically important algorithms operate complete continuously adaptable configurations. For example, Euclidean geometry assigns a ample domiciled to ruler-and-compass constructions, which manipulate geometric shapes. For immoderate shape, one tin find different that differs to an arbitrarily mini extent. Symbolic configurations manipulated by a Turing instrumentality do not differ to arbitrarily mini extent. Turing machines run complete discrete strings of elements (digits) drawn from a finite alphabet. One recurring contention concerns whether the integer paradigm is well-suited to exemplary intelligence activity aliases whether an analog paradigm would alternatively beryllium much fitting (MacLennan 2012; Piccinini and Bahar 2013).[1]

Besides introducing Turing machines, Turing (1936) proved several seminal mathematical results involving them. In particular, he proved the beingness of a universal Turing machine (UTM). Roughly speaking, a UTM is simply a Turing instrumentality that tin mimic immoderate different Turing machine. One provides the UTM pinch a symbolic input that codes the machine array for Turing instrumentality M. The UTM replicates M’s behavior, executing instructions enshrined by M’s instrumentality table. In that sense, the UTM is a programmable wide intent computer. To a first approximation, each individual computers are besides wide purpose: they can mimic immoderate Turing machine, erstwhile suitably programmed. The main caveat is that beingness computers person finite memory, whereas a Turing machine has unlimited memory. More accurately, then, a personal computer tin mimic immoderate Turing instrumentality until it exhausts its limited representation supply.

Turing’s chat helped laic the foundations for computer science, which seeks to design, build, and understand computing systems. As we know, machine scientists tin now build extremely sophisticated computing machines. All these machines implement something resembling Turing computation, though the specifications differ from Turing’s simplified model.

2. Artificial intelligence

Rapid advancement successful machine subject prompted many, including Turing, to contemplate whether we could build a machine tin of thought. Artificial intelligence (AI) intends to construct “thinking machinery”. More precisely, it intends to construct computing machines that execute halfway intelligence tasks specified arsenic reasoning, decision-making, problem solving, and truthful on. During the 1950s and 1960s, this extremity came to look progressively realistic (Haugeland 1985). A celebrated early occurrence was the Logic Theorist computer program (Newell and Simon 1956), which proved 38 of the first 52 theorems from Principia Mathematica (Whitehead and Russell 1925). In 1 case, it discovered a simpler impervious than Principia’s. Initial achievements of this kind stimulated tremendous liking wrong and extracurricular the academy. Many researchers predicted that intelligent machines were only a fewer years away. When assured predictions of reasoning machines proved too optimistic, galore observers mislaid liking aliases concluded that AI was a fool’s errand. Nevertheless, the decades person witnessed gradual progress, including immoderate striking caller advances. A few milestones:

• IBM’s Deep Blue defeated chess champion Gary Kasparov in 1997 (Campbell 1999).
• The driverless car Stanley completed a 132-mile people successful the Mojave Desert, winning the 2005 Defense Advanced Research Projects Agency (DARPA) Grand Challenge (Thrun, Montemerlo, Dahlkamp, et al. 2006).
• In 2012, AlexNet dramatically surpassed each erstwhile computational models successful a modular image classification task (Krizhevsky, Sutskever, and Hinton 2012).
• DeepMind’s AlphaGo defeated Lee Sedol, 1 of the apical Go players successful the world, successful 2016 (Silver, Schrittwieser, Simonyan, et al. 2017).
• In 2020, OpenAI released GPT-3, which generates uncannily human-like matter successful consequence to written prompts (Brown, Mann, Ryder, et al. 2020). An improved version, ChatGPT, was released successful 2022 and attracted wide societal attention.

These and different caller advances person sparked aggravated renewed focus upon AI, including galore commercialized applications.

Some philosophers insist that computers, nary matter really sophisticated they become, will astatine champion mimic alternatively than replicate thought. A machine simulation of the upwind does not really rain. A machine simulation of formation does not really fly. Even if a computing strategy could simulate intelligence activity, why suspect that it would represent the genuine article?

Turing (1950) anticipated these worries and tried to defuse them. He proposed a scenario, now called the Turing Test, wherever one evaluates whether an unseen interlocutor is simply a machine aliases a human. A computer passes the Turing test if 1 cannot find that it is simply a computer. Turing projected that we wantonness the question “Could a machine think?” arsenic hopelessly vague, replacing it pinch the mobility “Could a machine walk the Turing test?”. Turing’s chat has received considerable attention, proving particularly influential wrong AI. Ned Block (1981) offers an influential critique. He argues that definite possible machines walk the Turing trial moreover though these machines do not come close to genuine thought aliases intelligence. See the entry the Turing test for chat of Block’s objection and different issues surrounding the Turing Test. For chat of the Turing trial in relation to ChatGPT and akin models, spot (Bayne and Williams 2023; Floridi and Chiriatti 2020; Mahowald, Ivanova, Blank, et al. 2024).

For much connected AI, spot the entry logic and artificial intelligence. For overmuch much detail, spot Russell and Norvig (2022).

3. The classical computational mentation of mind

Warren McCulloch and Walter Pitts (1943) first suggested that something resembling the Turing instrumentality mightiness supply a bully exemplary for the mind. In the 1960s, Turing computation became cardinal to the emerging interdisciplinary inaugural cognitive science, which studies the mind by drafting upon psychology, machine science (especially AI), linguistics, philosophy, economics (especially game theory and behavioral economics), anthropology, and neuroscience. The label classical computational mentation of mind (which we will abbreviate arsenic CCTM) is now reasonably standard. According to CCTM, the mind is simply a computational strategy akin successful important respects to a Turing machine, and halfway intelligence processes (e.g., reasoning, decision-making, and problem solving) are computations akin in important respects to computations executed by a Turing machine. These formulations are imprecise. CCTM is champion seen arsenic a family of views, rather than a azygous well-defined view.[2]

It is communal to picture CCTM arsenic embodying “the computer metaphor”. This explanation is doubly misleading.

First, CCTM is amended formulated by describing the mind arsenic a “computing system” aliases a “computational system” rather than a “computer”. As David Chalmers (2011) notes, describing a strategy arsenic a “computer” powerfully suggests that the strategy is programmable. As Chalmers besides notes, 1 need not declare that the mind is programmable simply because 1 regards it as a Turing-style computational system. (Most Turing machines are not programmable.) Thus, the building “computer metaphor” strongly suggests theoretical commitments that are inessential to CCTM. The constituent present is not conscionable terminological. Critics of CCTM often object that the mind is not a programmable wide intent computer (Churchland, Koch, and Sejnowski 1990). Since classical computationalists request not declare (and usually do not claim) that the mind is simply a programmable wide intent computer, the objection is misdirected.

Second, CCTM is not intended metaphorically. CCTM does not simply hold that the mind is like a computing system. CCTM holds that the mind literally is a computing system. Of course, the most familiar artificial computing systems are made from silicon chips or similar materials, whereas the quality assemblage is made from soma and blood. But CCTM holds that this quality disguises a more fundamental similarity, which we tin seizure done a Turing-style computational model. In offering specified a model, we prescind from physical details. We attain an absurd computational explanation that could beryllium physically implemented successful divers ways (e.g., done silicon chips, aliases neurons, aliases pulleys and levers). CCTM holds that a suitable abstract computational exemplary offers a virtually existent explanation of core intelligence processes.

It is communal to summarize CCTM done the motto “the mind is a Turing machine”. This motto is besides somewhat misleading, because nary 1 regards Turing’s precise formalism arsenic a plausible model of intelligence activity. The formalism seems excessively restrictive in several ways:

• Turing machines execute axenic symbolic computation. The inputs and outputs are symbols inscribed successful representation locations. In contrast, the mind receives sensory input (e.g., retinal stimulations) and produces motor output (e.g., musculus activations). A complete theory must picture really intelligence computation interfaces pinch sensory inputs and centrifugal outputs.
• A Turing instrumentality has infinite discrete representation capacity. Ordinary biological systems person finite representation capacity. A plausible psychological exemplary must switch the infinite representation shop pinch a large but finite representation store
• Modern computers person random entree memory: addressable memory locations that the cardinal processor tin straight access. Turing instrumentality representation is not addressable. The cardinal processor can access a location only by sequentially accessing intermediate locations. Computation without addressable representation is hopelessly inefficient. For that reason, C.R. Gallistel and Adam King (2009) argue that addressable representation gives a amended exemplary of the mind than non-addressable memory.
• A Turing instrumentality has a cardinal processor that operates serially, executing 1 instruction astatine a time. Other computational formalisms relax this assumption, allowing multiple processing units that run successful parallel. Classical computationalists tin let parallel computations (Fodor and Pylyshyn 1988; Gallistel and King 2009: 174). See Gandy (1980) and Sieg (2009) for wide mathematical treatments that encompass some serial and parallel computation.
• Turing computation is deterministic: full computational state determines consequent computational state. One mightiness instead allow stochastic computations. In a stochastic model, current state does not dictate a unsocial adjacent state. Rather, location is simply a certain probability that the instrumentality will modulation from 1 authorities to another.

CCTM claims that intelligence activity is “Turing-style computation”, allowing these and different departures from Turing’s ain formalism.

3.1 Machine functionalism

Hilary Putnam (1967) introduced CCTM into philosophy. He contrasted his position pinch logical behaviorism and type-identity theory. Each position purports to uncover the quality of mental states, including propositional attitudes (e.g., beliefs), sensations (e.g., pains), and emotions (e.g., fear). According to logical behaviorism, intelligence states are behavioral dispositions. According to type-identity theory, intelligence states are encephalon states. Putnam advances an opposing functionalist view, connected which intelligence states are functional states. According to functionalism, a strategy has a mind when the strategy has a suitable functional organization. Mental states are states that play due roles successful the system’s functional organization. Each intelligence authorities is individuated by its interactions pinch sensory input, centrifugal output, and other intelligence states.

Functionalism offers notable advantages complete logical behaviorism and type-identity theory:

• Behaviorists want to subordinate each intelligence authorities pinch a characteristic shape of behavior—a hopeless task, because individual intelligence states do not usually person characteristic behavioral effects. Behavior almost ever results from chopped intelligence states operating together (e.g., a belief and a desire). Functionalism avoids this trouble by individuating intelligence states done characteristic relations not only to sensory input and behaviour but besides to one another.
• Type-identity theorists want to subordinate each intelligence authorities pinch a characteristic beingness aliases neurophysiological state. Putnam casts this project into uncertainty by arguing that intelligence states are multiply realizable: the aforesaid intelligence authorities tin beryllium realized by diverse physical systems, including not only terrestrial creatures but also hypothetical creatures (e.g., a silicon-based Martian). Functionalism is tailor-made to accommodate aggregate realizability. According to functionalism, what matters for mentality is simply a shape of organization, which could beryllium physically realized successful galore different ways. See the entry multiple realizability for further chat of this argument.

Putnam defends a marque of functionalism now called machine functionalism. He emphasizes probabilistic automata, which are akin to Turing machines isolated from that transitions between computational states are stochastic. He proposes that intelligence activity implements a probabilistic automaton and that peculiar intelligence states are instrumentality states of the automaton’s cardinal processor. The machine array specifies an due functional organization, and it also specifies the domiciled that individual intelligence states play wrong that functional organization. In this way, Putnam combines functionalism with CCTM.

Machine functionalism faces respective problems. One problem, highlighted by Ned Block and Jerry Fodor (1972), concerns the productivity of thought. A normal quality tin entertain a imaginable infinity of propositions. Machine functionalism identifies intelligence states with machine states of a probabilistic automaton. Since location are only finitely galore instrumentality states, location are not capable instrumentality states to pair one-one pinch imaginable intelligence states of a normal human. Of course, an existent quality will only ever entertain finitely galore propositions. However, Block and Fodor contend that this limitation reflects limits on lifespan and memory, alternatively than (say) immoderate psychological rule that restricts the people of humanly entertainable propositions. A probabilistic automaton is endowed pinch unlimited clip and memory capacity yet moreover still has only finitely galore instrumentality states. Apparently, then, instrumentality functionalism mislocates the finitary limits upon quality cognition.

Another problem for instrumentality functionalism, besides highlighted by Block and Fodor (1972), concerns the systematicity of thought. An ability to entertain 1 proposition is correlated pinch an expertise to think different propositions. For example, personification who tin entertain the thought that John loves Mary tin besides entertain the thought that Mary loves John. Thus, location look to beryllium systematic relations betwixt intelligence states. A bully mentation should bespeak those systematic relations. Yet instrumentality functionalism identifies mental states pinch unstructured machines states, which deficiency the requisite systematic relations to another. For that reason, machine functionalism does not explicate systematicity. In consequence to this objection, instrumentality functionalists mightiness contradict that they are obligated to explicate systematicity. Nevertheless, the objection suggests that machine functionalism neglects basal features of quality mentality. A amended mentation would explicate those features successful a opinionated way.

While the productivity and systematicity objections to machine functionalism are possibly not decisive, they supply beardown impetus to pursue an improved type of CCTM. See Block (1978) for additional problems facing instrumentality functionalism and functionalism more generally.

3.2 The representational mentation of mind

Fodor (1975, 1981, 1987, 1990, 1994, 2008) advocates a type of CCTM that accommodates systematicity and productivity overmuch more satisfactorily. He shifts attraction to the symbols manipulated during Turing-style computation.

An aged view, stretching backmost astatine slightest to William of Ockham’s Summa Logicae, holds that reasoning occurs successful a language of thought (sometimes called Mentalese). Fodor revives this view. He postulates a strategy of intelligence representations, including both primitive representations and analyzable representations formed from primitive representations. For example, the primitive Mentalese words JOHN, MARY, and LOVES tin harvester to shape the Mentalese condemnation JOHN LOVES MARY. Mentalese is compositional: the meaning of a complex Mentalese look is simply a usability of the meanings of its parts and the measurement those parts are combined. Propositional attitudes are relations to Mentalese symbols. Fodor calls this position the representational mentation of mind (RTM). Combining RTM with CCTM, he argues that intelligence activity involves Turing-style computation complete the connection of thought. Mental computation stores Mentalese symbols successful representation locations, manipulating those symbols in accord pinch mechanical rules.

A premier kindness of RTM is really readily it accommodates productivity and systematicity:

Productivity: RTM postulates a finite group of primitive Mentalese expressions, combinable into a imaginable infinity of complex Mentalese expressions. A philosopher pinch entree to primitive Mentalese vocabulary and Mentalese compounding devices has the imaginable to entertain an infinity of Mentalese expressions. She truthful has the potential to instantiate infinitely galore propositional attitudes (neglecting limits connected clip and memory).

Systematicity: According to RTM, location are systematic relations betwixt which propositional attitudes a philosopher can entertain. For example, suppose I tin deliberation that John loves Mary. According to RTM, my doing truthful involves my opinionated successful immoderate relation R to a Mentalese condemnation JOHN LOVES MARY, composed of Mentalese words JOHN, LOVES, and MARY mixed successful the correct way. If I have this capacity, past I besides person the capacity to guidelines successful relation R to the chopped Mentalese condemnation MARY LOVES JOHN, thereby thinking that Mary loves John. So the capacity to deliberation that John loves Mary is systematically related to the capacity to deliberation that Mary loves John.

By treating propositional attitudes arsenic relations to analyzable mental symbols, RTM explains some productivity and systematicity.

CCTM+RTM differs from instrumentality functionalism successful respective different respects. First, instrumentality functionalism is simply a mentation of intelligence states in general, while RTM is only a mentation of propositional attitudes. Second, proponents of CCTM+RTM request not opportunity that propositional attitudes are individuated functionally. As Fodor (2000: 105, fn. 4) notes, we must separate computationalism (mental processes are computational) from functionalism (mental states are functional states). Machine functionalism endorses some doctrines. CCTM+RTM endorses only the first. Unfortunately, galore philosophers still mistakenly presume that computationalism entails a functionalist approach to propositional attitudes (see Piccinini 2004 for discussion).

Philosophical chat of RTM tends to attraction chiefly on high-level quality thought, particularly belief and desire. However, CCTM+RTM is applicable to a overmuch wider scope of intelligence states and processes. Many cognitive scientists use it to non-human animals. For example, Gallistel and King (2009) use it to certain invertebrate phenomena (e.g., honeybee navigation). Even confining attention to humans, 1 tin use CCTM+RTM to subpersonal processing. Fodor (1983) argues that cognition involves a subpersonal “module” that converts retinal input into Mentalese symbols and past performs computations complete those symbols. Thus, talk astir a connection of thought is potentially misleading, since it suggests a non-existent regularisation to higher-level intelligence activity.

Also perchance misleading is the explanation of Mentalese arsenic a language, which suggests that each Mentalese symbols resemble expressions successful a earthy language. Many philosophers, including Fodor, sometimes look to endorse that position. However, location are possible non-propositional formats for intelligence representations. Proponents of CCTM+RTM tin adopt a pluralistic line, allowing intelligence computation to operate complete items akin to images, maps, diagrams, aliases other non-propositional representations (Johnson-Laird 2004: 187; McDermott 2001: 69; Pinker 2005: 7; Sloman 1978: 144–176). The pluralistic line seems particularly plausible arsenic applied to subpersonal processes (such arsenic perception) and non-human animals. Michael Rescorla (2009a, 2009b) surveys investigation connected cognitive maps (Tolman 1948; O’Keefe and Nadel 1978; Gallistel 1990), suggesting that some animals whitethorn navigate by computing complete intelligence representations more similar to maps than sentences. Elisabeth Camp (2009), citing research on baboon societal relationship (Cheney and Seyfarth 2007), argues that baboons whitethorn encode societal power relations done non-sentential tree-structured representations.

CCTM+RTM is schematic. To capable successful the schema, 1 must provide detailed computational models of circumstantial intelligence processes. A complete model will:

• describe the intelligence representations manipulated by the process;
• isolate simple operations that manipulate the representations (e.g., inscribing a awesome successful a representation location); and
• delineate mechanical rules governing exertion of elementary operations.

By providing a elaborate computational model, we decompose a complex mental process into a bid of simple operations governed by precise, regular instructions.

CCTM+RTM remains neutral successful the accepted statement betwixt physicalism and constituent dualism. A Turing-style exemplary proceeds astatine a very abstract level, not saying whether intelligence computations are implemented by beingness worldly aliases Cartesian soul-stuff (Block 1983: 522). In practice, each proponents of CCTM+RTM clasp a broadly physicalist outlook. They clasp that intelligence computations are implemented not by soul-stuff but alternatively by the brain. On this view, mental representations are realized by neural states, and computational operations complete intelligence representations are realized by neural processes. Ultimately, physicalist proponents of CCTM+RTM must produce empirically well-confirmed theories that explicate really precisely neural activity implements Turing-style computation. As Gallistel and King (2009) emphasize, we do not presently person specified theories—though see Zylberberg, Dehaene, Roelfsema, and Sigman (2011) and Akhlaghpour (2022) for immoderate speculations.

Fodor (1975) advances CCTM+RTM arsenic a instauration for cognitive science. He discusses intelligence phenomena specified arsenic decision-making, perception, and linguistic processing. In each case, he maintains, our champion scientific theories postulate Turing-style computation complete mental representations. In fact, he argues that our only viable theories person this form. He concludes that CCTM+RTM is “the only game successful town”. Many cognitive scientists reason on similar lines. C.R. Gallistel and Adam King (2009), Philip Johnson-Laird (1988), Allen Newell and Herbert Simon (1976), and Zenon Pylyshyn (1984) each urge Turing-style computation complete intelligence symbols as the champion instauration for technological theorizing astir the mind.

4. Neural networks

In the 1980s, connectionism emerged arsenic a salient rival to classical computationalism. Connectionists tie inspiration from neurophysiology rather than logic and machine science. They employment computational models, neural networks, that disagree importantly from Turing-style models. A neural network is simply a postulation of interconnected nodes. Nodes autumn into 3 categories: input nodes, output nodes, and hidden nodes (which mediate between input and output nodes). Nodes person activation values, given by existent numbers. One node tin carnivore a weighted connection to another node, besides fixed by a existent number. Activations of input nodes are wished exogenously: these are the inputs to computation. Total input activation of a hidden aliases output node is a weighted sum of the activations of nodes feeding into it. Activation of a hidden aliases output node is simply a usability of its full input activation; the peculiar usability varies pinch the network. During neural web computation, waves of activation propagate from input nodes to output nodes, arsenic wished by weighted connections between nodes.

In a feedforward network, weighted connections travel only in one direction. Recurrent networks person feedback loops, in which connections emanating from hidden units circle backmost to hidden units. Recurrent networks are little mathematically tractable than feedforward networks. However, they fig crucially successful psychological modeling of various phenomena, specified arsenic phenomena that impact some kind of representation (Elman 1990).

Weights successful a neural web are typically mutable, evolving successful accord with a learning algorithm. The lit offers various learning algorithms, but the basal thought is usually to set weights so that actual outputs gradually move person to the target outputs 1 would expect for the applicable inputs. The backpropagation algorithm is simply a wide utilized algorithm of this kind (Rumelhart, Hinton, and Williams 1986).

Connectionism traces backmost to McCulloch and Pitts (1943), who studied networks of interconnected logic gates (e.g., AND-gates and OR-gates). One tin position a web of logic gates arsenic a neural network, with activations confined to 2 values (0 and 1) and activation functions fixed by the accustomed truth-functions. McCulloch and Pitts advanced logic gates arsenic idealized models of individual neurons. Their discussion exerted a profound power connected machine subject (von Neumann 1945). Modern integer computers are simply networks of logic gates. Within cognitive science, however, researchers usually focus upon networks whose elements are much “neuron-like” than logic gates. In particular, modern-day connectionists typically emphasize analog neural networks whose nodes return continuous rather than discrete activation values. Some authors moreover usage the phrase “neural network” truthful that it exclusively denotes such networks.

Neural networks received comparatively scant attraction from cognitive scientists during the 1960s and 1970s, erstwhile Turing-style models dominated. The 1980s witnessed a immense resurgence of liking successful neural networks, particularly analog neural networks, pinch the two-volume Parallel Distributed Processing (Rumelhart, McClelland, and the PDP investigation group, 1986; McClelland, Rumelhart, and the PDP research group, 1987) serving arsenic a manifesto. Researchers constructed connectionist models of divers phenomena: entity recognition, speech perception, condemnation comprehension, cognitive development, and truthful on. Impressed by connectionism, galore researchers concluded that CCTM+RTM was nary longer “the only crippled successful town”.

In the 2010s, a people of computational models known arsenic deep neural networks became rather celebrated (Krizhevsky, Sutskever, and Hinton 2012; LeCun, Bengio, and Hinton 2015). These models are neural networks pinch aggregate layers of hidden nodes (sometimes hundreds of such layers). Deep neural networks—trained connected ample information sets through 1 aliases different learning algorithm (usually backpropagation)—have achieved awesome occurrence successful galore areas of AI, including image classification (AlexNet), strategical game-playing (AlphaGo), and earthy connection processing (ChatGPT). Deep neural networks are now wide deployed successful commercialized applications, and they are the attraction of extended ongoing investigation wrong some academia and industry. Researchers person besides utilized them to exemplary the mind (e.g. Kriegeskorte 2015; Marblestone, Wayne, and Kording 2016; Storrs, Kietzmann, Walther, et al. 2021; Zhuang, Yan, Nayebi, et al. 2021), although really overmuch occurrence this modeling endeavor has frankincense far achieved is simply a arguable mobility (Bowers, Malholtra, Dujmović, et al. 2023).

For a elaborate overview of neural networks, spot Haykin (2008). For a user-friendly introduction, pinch an accent connected psychological applications, spot Marcus (2001). For a philosophically oriented introduction to heavy neural networks, spot Buckner (2019). For connections betwixt heavy neural networks and the history of philosophy, spot Buckner (2024).

4.1 Relation betwixt neural networks and classical computation

Neural networks person a very different “feel” than classical (i.e., Turing-style) models. Yet classical computation and neural web computation are not mutually exclusive:

• One tin instrumentality a neural web successful a classical model. Indeed, each neural web ever physically constructed has been implemented connected a integer computer.
• One tin instrumentality a classical exemplary successful a neural network. Modern integer computers instrumentality Turing-style computation in networks of logic gates. Alternatively, 1 tin instrumentality Turing-style computation utilizing an analog recurrent neural web whose nodes take continuous activation values (Graves, Wayne, and Danihelka 2014, Other Internet Resources; Siegelmann and Sontag 1991; Siegelmann and Sontag 1995).

Although immoderate researchers propose a basal guidance between classical computation and neural web computation, it seems more accurate to place 2 modeling traditions that overlap successful certain cases but not others (cf. Boden 1991; Piccinini 2008b). In this connection, it is besides worthy noting that classical computationalism and connectionist computationalism person their communal root successful the work of McCulloch and Pitts.

Philosophers often opportunity that classical computation involves “rule-governed awesome manipulation” while neural network computation is non-symbolic. The intuitive image is that “information” successful neural networks is globally distributed across the weights and activations, alternatively than concentrated in localized symbols. However, the conception of “symbol” itself requires explication, truthful it is often unclear what theorists mean by describing computation arsenic symbolic versus non-symbolic. As mentioned in §1, the Turing formalism places very fewer conditions on “symbols”. Regarding primitive symbols, Turing assumes just that location are finitely galore of them and that they tin be inscribed successful read/write representation locations. Neural networks tin also manipulate symbols satisfying these 2 conditions: arsenic conscionable noted, one can instrumentality a Turing-style exemplary successful a neural network.

Many discussions of the symbolic/non-symbolic dichotomy employment a more robust conception of “symbol”. On the much robust approach, a symbol is the benignant of point that represents a taxable matter. Thus, something is simply a awesome only if it has semantic aliases representational properties. If we employment this much robust conception of symbol, past the symbolic/non-symbolic favoritism cross-cuts the favoritism between Turing-style computation and neural web computation. A Turing machine request not employment symbols successful the much robust sense. As acold as the Turing formalism goes, symbols manipulated during Turing computation request not person representational properties (Chalmers 2011). Conversely, a neural web tin manipulate symbols with representational properties. Indeed, an analog neural web can manipulate symbols that person a combinatorial syntax and semantics (Horgan and Tienson 1996; Marcus 2001).

Following Steven Pinker and Alan Prince (1988), we whitethorn distinguish between eliminative connectionism and implementationist connectionism.

Eliminative connectionists beforehand connectionism arsenic a rival to classical computationalism. They reason that the Turing formalism is irrelevant to psychological explanation. Often, though not always, they activity to revive the associationist contented in psychology, a contented that CCTM had forcefully challenged. Often, though not always, they onslaught the mentalist, nativist linguistics pioneered by Noam Chomsky (1965). Often, though not always, they manifest overt hostility to the very conception of intelligence representation. But the defining characteristic of eliminative connectionism is that it uses neural networks arsenic replacements for Turing-style models. Eliminative connectionists position the mind arsenic a computing strategy of a radically different benignant than the Turing machine. A fewer authors explicitly espouse eliminative connectionism (Churchland 1989; Rumelhart and McClelland 1986; Horgan and Tienson 1996), and many others incline towards it.

Implementationist connectionism is simply a much ecumenical position. It allows a perchance valuable domiciled for some Turing-style models and neural networks, operating harmoniously astatine different levels of explanation (Marcus 2001; Smolensky 1988). A Turing-style model is higher-level, whereas a neural web exemplary is lower-level. The neural web illuminates really the encephalon implements the Turing-style model, conscionable arsenic a explanation successful position of logic gates illuminates really a individual machine executes a programme successful a high-level programming language.

4.2 Arguments for connectionism

Connectionism excites galore researchers because of the affinity between neural networks and the brain. Nodes lucifer neurons, while connections betwixt nodes lucifer synapses. Connectionist modeling therefore seems much “biologically plausible” than classical modeling. A connectionist exemplary of a psychological phenomenon apparently captures (in an idealized way) how interconnected neurons mightiness make the phenomenon.

When evaluating the statement from biologic plausibility, 1 should recognize that neural networks alteration wide successful really intimately they match actual encephalon activity. Many networks that fig prominently in connectionist writings are not truthful biologically plausible (Bechtel and Abrahamsen 2002: 341–343; Bermúdez 2010: 237–239; Clark 2014: 87–89; Harnish 2002: 359–362). A few examples:

• Real neurons are overmuch much heterogeneous than the interchangeable nodes that fig successful emblematic connectionist networks.
• Real neurons emit discrete spikes (action potentials) arsenic outputs. But the nodes that fig successful galore salient neural networks, including the champion known heavy neural networks, alternatively person continuous outputs.
• The backpropagation algorithm requires that weights betwixt nodes can alteration betwixt excitatory and inhibitory, yet existent synapses cannot so alteration (Crick and Asanuma 1986). Moreover, accepted applications of the algorithm presume target outputs supplied exogenously by modelers who cognize the desired answer. In that sense, learning is supervised. Very small learning successful existent biological systems involves thing resembling supervised training.

On the different hand, immoderate neural networks are much biologically plausible (Buckner and Garson 2019; Illing, Gerstner, and Brea 2019). For example, location are neural networks whose nodes output discrete spikes astir akin to those emitted by existent neurons successful the brain (Maass 1996; Buesing, Bill, Nessler, and Maass 2011). Furthermore, a large lit seeks to articulate biologically realistic connectionist learning algorithms, sometimes by approximating backpropagation (e.g. Lillicrap et al. 2016; Whittington and Bogacz, 2017), sometimes by replacing it pinch an replacement attack (e.g. Krotov and Hopfield 2019). Lillicrap et al. (2020) reason astatine length that backpropagation tin beryllium developed successful a biologically plausible way. They statement successful peculiar that, though backpropagation was traditionally mixed pinch supervised learning, it tin alternatively be combined pinch unsupervised learning (e.g. Kingma and Welling 2019) or with reinforcement learning (e.g. Silver et al. 2016).

Even erstwhile a neural web is not biologically plausible, it whitethorn still be more biologically plausible than classical models. Neural networks surely look person than Turing-style models, successful both details and spirit, to neurophysiological description. Many cognitive scientists interest that CCTM reflects a misguided effort astatine imposing the architecture of integer computers onto the brain. Some uncertainty that the encephalon implements thing resembling integer computation, i.e., computation complete discrete configurations of digits (Piccinini and Bahar 2013). Others uncertainty that brains show cleanable Turing-style separation betwixt cardinal processor and read/write representation (Dayan 2009). Neural networks fare amended connected some scores: they do not require computation complete discrete configurations of digits, and they do not postulate a cleanable separation betwixt cardinal processor and read/write memory.

Classical computationalists typically reply that it is premature to draw patient conclusions based upon biologic plausibility, fixed how little we understand astir the narration betwixt neural, computational, and cognitive levels of explanation (Gallistel and King 2009; Marcus 2001). Using measurement techniques specified arsenic compartment recordings and functional magnetic resonance imaging (fMRI), and drafting upon disciplines arsenic divers arsenic physics, biology, AI, accusation theory, statistics, chart theory, and dynamical systems theory, neuroscientists person accumulated important knowledge astir the brain at varying levels of granularity (Zednik 2019). We now cognize rather a lot astir individual neurons, astir really neurons interact wrong neural populations, astir the localization of intelligence activity successful cortical regions (e.g. the ocular cortex), and astir interactions among cortical regions. Yet we still person a tremendous magnitude to study about how neural insubstantial accomplishes the tasks that it surely accomplishes: perception, reasoning, decision-making, connection acquisition, and so on. Given our coming authorities of comparative ignorance, it would beryllium rash to insist that the encephalon does not instrumentality thing resembling Turing computation.

Connectionists connection galore further arguments that we should employ connectionist models alternatively of, aliases successful summation to, classical models. See the entry connectionism for an overview. For purposes of this entry, we mention two additional arguments.

The first statement emphasizes learning (Bechtel and Abrahamsen 2002: 51). A immense scope of cognitive phenomena involve learning from experience. Many connectionist models are explicitly designed to exemplary learning, done backpropagation aliases immoderate other algorithm that modifies the weights betwixt nodes. By contrast, connectionists often kick that location are nary bully classical models of learning. Classical computationalists tin respond by citing perceived defects of connectionist learning algorithms. Classical computationalists tin besides mention Bayesian determination theory, a mathematical exemplary of conclusion and decision-making nether uncertainty. In the Bayesian framework, uncertainty is codified through probability. Precise rules dictate really to update probabilities successful ray of caller grounds and really to prime actions in light of probabilities and utilities. (See the entries Bayes’s theorem and normative theories of logical choice: expected utility for details.) Bayesian cognitive science uses Bayesian decision mentation to conception mathematical models of intelligence activity (Ma 2019; Ma, Kording, and Goldreich 2023). Over the past fewer decades, Bayesian cognitive subject has accrued galore explanatory successes. This awesome way grounds suggests that immoderate intelligence processes are Bayesian aliases astir Bayesian (Rescorla forthcoming). Moreover, classical computing systems tin execute aliases astatine slightest approximately execute Bayesian updating successful various realistic scenarios (Murphy 2023; Thrun, Burgard, and Fox 2005). Arguably, then, classical computation can exemplary galore important cases of learning.

The 2nd statement emphasizes speed of computation. Neurons are overmuch slower than silicon-based components of integer computers. For this reason, neurons could not execute serial computation quickly enough to lucifer accelerated quality capacity successful perception, linguistic comprehension, decision-making, etc. Connectionists support that the only viable solution is to switch serial computation pinch a “massively parallel” computational architecture—precisely what neural networks supply (Feldman and Ballard 1982; Rumelhart 1989). However, this statement is only effective against classical computationalists who insist upon serial processing. As noted in §3, immoderate Turing-style models impact parallel processing. Many classical computationalists are happy to let “massively parallel” mental computation, and the statement gains nary traction against these researchers. That being said, the statement highlights an important question that immoderate computationalist—whether classical, connectionist, aliases otherwise—must address: How does a encephalon built from comparatively slow neurons execute blase computations so quickly? Neither classical nor connectionist computationalists have answered this mobility satisfactorily (Gallistel and King 2009: 174 and 265).

4.3 Systematicity and productivity

Fodor and Pylyshyn (1988) connection a wide discussed critique of eliminativist connectionism. They reason that systematicity and productivity neglect successful connectionist models, isolated from erstwhile the connectionist exemplary implements a classical model. Hence, connectionism does not furnish a viable replacement to CCTM. At best, it supplies a low-level explanation that helps span the spread betwixt Turing-style computation and neuroscientific description.

This statement has elicited galore replies and counter-replies. Some argue that neural networks tin grounds systematicity without implementing thing for illustration classical computational architecture (Horgan and Tienson 1996; Chalmers 1990; Smolensky 1991; van Gelder 1990). Some reason that Fodor and Pylyshyn vastly exaggerate systematicity (Johnson 2004) aliases productivity (Rumelhart and McClelland 1986), particularly for non-human animals (Dennett 1991). These issues, and galore others raised by Fodor and Pylyshyn’s argument, have been thoroughly investigated successful the literature. For further discussion, spot Bechtel and Abrahamsen (2002: 156–199), Bermúdez (2005: 244–278), Chalmers (1993), Clark (2014: 84–86), and the encyclopedia entries on the connection of thought hypothesis and on connectionism.

Gallistel and King (2009) beforehand a related but chopped productivity argument. They stress productivity of intelligence computation, as opposed to productivity of intelligence states. Through detailed empirical lawsuit studies, they reason that galore non-human animals can extract, store, and retrieve elaborate records of the surrounding environment. For example, the Western scrub jay records wherever it cached food, what benignant of nutrient it cached successful each location, erstwhile it cached the food, and whether it has depleted a fixed cache (Clayton, Emery, and Dickinson 2006). The jay tin entree these records and exploit them successful divers computations: computing whether a nutrient item stored successful immoderate cache is apt to person decayed; computing a way from one location to another; and truthful on. The number of possible computations a jay tin execute is, for each applicable purposes, infinite.

CCTM explains the productivity of intelligence computation by positing a central processor that stores and retrieves symbols successful addressable read/write memory. When needed, the cardinal processor tin retrieve arbitrary, unpredicted combinations of symbols from memory. In contrast, Gallistel and King argue, connectionism has difficulty accommodating the productivity of intelligence computation. Although Gallistel and King do not cautiously separate betwixt eliminativist and implementationist connectionism, we whitethorn summarize their argument as follows:

• Eliminativist connectionism cannot explicate really organisms combine stored memories (e.g., cache locations) for computational purposes (e.g., computing a way from 1 cache to another). There are a virtual infinity of imaginable combinations that mightiness beryllium useful, with no predicting successful beforehand which pieces of accusation must beryllium combined in early computations. The only computationally tractable solution is symbol retention successful readily accessible read/write memory locations—a solution that eliminativist connectionists reject.
• Implementationist connectionists tin postulate awesome retention in read/write memory, as implemented by a neural network. However, the mechanisms that connectionists usually propose for implementing representation are not plausible. Existing proposals are mainly variants upon a azygous idea: a recurrent neural web that allows reverberating activity to recreation astir a loop (Elman 1990). There are many reasons why the reverberatory loop exemplary is hopeless arsenic a theory of semipermanent memory. For example, sound successful the tense strategy ensures that signals would quickly degrade successful a fewer minutes. Implementationist connectionists person frankincense acold offered nary plausible exemplary of read/write memory.[3]

Gallistel and King reason that CCTM is overmuch amended suited than either eliminativist aliases implementationist connectionism to explicate a vast scope of cognitive phenomena.

Critics onslaught this caller productivity statement from various angles, focusing chiefly connected the empirical lawsuit studies adduced by Gallistel and King. Peter Dayan (2009), John Donahoe (2010), and Christopher Mole (2014) reason that biologically plausible neural web models can accommodate astatine slightest immoderate of the lawsuit studies. Dayan and Donahoe argue that empirically capable neural web models tin dispense with anything resembling read/write memory. Mole argues that, successful certain cases, empirically capable neural web models can implement the read/write representation mechanisms posited by Gallistel and King. Debate connected these basal issues seems poised to continue good into the future.

4.4 Computational neuroscience

Computational neuroscience describes the tense system through computational models (Trappenberg 2010; Miller 2018). Although computational neuroscience is grounded successful mathematical modeling of individual neurons, its unique attraction is systems of interconnected neurons. Computational neuroscientists typically model these systems arsenic neural networks. This investigation whitethorn beryllium seen arsenic a variant, off-shoot, aliases descendant of connectionism. However, most computational neuroscientists do not self-identify arsenic connectionists. There are respective differences betwixt connectionism and computational neuroscience:

• Neural networks employed by computational neuroscientists are much more biologically realistic than those employed by connectionists. The computational neuroscience lit is filled pinch talk astir firing rates, action potentials, tuning curves, etc. These notions play at best a constricted domiciled successful connectionist research, specified arsenic astir of the research canvassed successful (Rogers and McClelland 2014).
• Computational neuroscience is driven successful ample measurement by knowledge about the brain, and it assigns immense value to neurophysiological data (e.g., compartment recordings). Connectionists spot overmuch little emphasis upon specified data. Their investigation is chiefly driven by behavioral data (although much caller connectionist writings mention neurophysiological data pinch somewhat greater frequency).
• Computational neuroscientists usually respect individual nodes in neural networks arsenic idealized descriptions of existent neurons. Connectionists usually alternatively respect nodes arsenic neuron-like processing units (Rogers and McClelland 2014) while remaining neutral astir really precisely these units representation onto actual neurophysiological entities.

One mightiness opportunity that computational neuroscience is concerned chiefly with neural computation (computation by systems of neurons), whereas connectionism is concerned chiefly pinch absurd computational models inspired by neural computation. But the boundaries between connectionism and computational neuroscience are admittedly somewhat porous. Doerig, Sommers, Seeliger, et al. (2023) propose the label neuroconnectionism for a investigation programme that thoroughly integrates neuroscience pinch neural web modeling.

Serious philosophical engagement pinch neuroscience dates backmost astatine least to Patricia Churchland’s Neurophilosophy (1986). As computational neuroscience matured, Churchland became 1 of its main philosophical champions (Churchland, Koch, and Sejnowski 1990; Churchland and Sejnowski 1992). She was joined by Paul Churchland (1995, 2007) and others (Eliasmith 2013; Eliasmith and Anderson 2003; Piccinini and Bahar 2013; Piccinini and Shagrir 2014). All these authors clasp that theorizing astir intelligence computation should begin with the brain, not pinch Turing machines aliases different inappropriate tools drawn from logic and machine science. They besides clasp that neural network modeling should strive for greater biologic realism than connectionist models typically attain. Chris Eliasmith (2013) develops this neurocomputational viewpoint done the Neural Engineering Framework, which supplements computational neuroscience with tools drawn from power mentation (Brogan 1990). He intends to “reverse engineer” the brain, building large-scale, biologically plausible neural web models of cognitive phenomena.

Computational neuroscience differs successful a important respect from CCTM and connectionism: it abandons aggregate realizability. Computational neuroscientists mention circumstantial neurophysiological properties and processes, truthful their models do not use arsenic good to (say) a sufficiently different silicon-based creature. Thus, computational neuroscience sacrifices a cardinal characteristic that primitively attracted philosophers to CTM. Computational neuroscientists will respond that this sacrifice is worthy the resultant penetration into neurophysiological underpinnings. But galore computationalists interest that, by focusing too much connected neural underpinnings, we consequence losing show of the cognitive forest for the neuronal trees. Neurophysiological specifications are important, but don’t we besides request an further absurd level of computational explanation that prescinds from specified details? Gallistel and King (2009) reason that a myopic fixation upon what we currently cognize astir the encephalon has led computational neuroscience to shortchange halfway cognitive phenomena specified arsenic navigation, spatial and temporal learning, and truthful on. Similarly, Edelman (2014) complains that the Neural Engineering Framework substitutes a blizzard of neurophysiological specifications for satisfying psychological explanations.

Partly successful consequence to specified worries, immoderate researchers propose an integrated cognitive computational neuroscience that connects psychological theories pinch neural implementation mechanisms (Naselaris et al. 2018; Kriegeskorte and Douglas 2018). The basal idea is to usage neural web models to illuminate really intelligence processes are instantiated successful the brain, thereby grounding multiply realizable cognitive explanation successful the neurophysiological. A bully illustration is recent activity connected neural implementation of Bayesian conclusion (Pouget et al. 2013; Orhan and Ma 2017; Aitchison and Lengyel 2016). Researchers articulate (multiply realizable) Bayesian models of various mental processes; they conception biologically plausible neural networks that execute aliases astir execute the posited Bayesian computations; and they measure really good these neural web models fresh with neurophysiological data.

Despite the differences betwixt connectionism and computational neuroscience, these 2 movements raise galore akin issues. In particular, the dialectic from §4.4 regarding systematicity and productivity arises successful akin form.

5. Computation and representation

Philosophers and cognitive scientists usage the term “representation” successful divers ways. Within philosophy, the most ascendant usage ties practice to intentionality, i.e., the “aboutness” of intelligence states. Contemporary philosophers usually elucidate intentionality by invoking representational content. A representational intelligence authorities has a contented that represents the world arsenic being a definite way, truthful we tin inquire whether the world is so that way. Thus, representationally contentful mental states are semantically evaluable pinch respect to properties such arsenic truth, accuracy, fulfillment, and truthful on. To illustrate:

• Beliefs are the sorts of things that tin beryllium existent aliases false. My belief that Emmanuel Macron is French is existent if Emmanuel Macron is French, mendacious if he is not.
• Perceptual states are the sorts of things that tin beryllium meticulous or inaccurate. My perceptual acquisition as of a reddish sphere is accurate only if a reddish sphere is earlier me.
• Desires are the sorts of things that tin fulfilled aliases thwarted. My desire to eat chocolate is fulfilled if I eat chocolate, thwarted if I do not eat chocolate.

Beliefs person truth-conditions (conditions nether which they are true), perceptual states person accuracy-conditions (conditions nether which they are accurate), and desires person fulfillment-conditions (conditions nether which they are fulfilled).

In mean life, we often foretell and explicate behaviour by invoking beliefs, desires, and different representationally contentful mental states. We place these states done their representational properties. When we opportunity “Frank believes that Emmanuel Macron is French”, we specify the information nether which Frank’s belief is existent (namely, that Emmanuel Macron is French). When we say “Frank wants to eat chocolate”, we specify the condition under which Frank’s desire is fulfilled (namely, that Frank eats chocolate). So people psychology assigns a cardinal domiciled to intentional descriptions, i.e., descriptions that identify mental states done their representational properties. Whether scientific psychology should likewise employment intentional descriptions is a contested rumor wrong modern accuracy of mind.

Intentional realism is realism regarding representation. At a minimum, this position holds that representational properties are genuine aspects of mentality. Usually, it is besides taken to clasp that scientific psychology should freely employment intentional descriptions when appropriate. Intentional realism is simply a celebrated position, advocated by Tyler Burge (2010a), Jerry Fodor (1987), Christopher Peacocke (1992, 1994), and galore others. One salient statement for intentional realism cites cognitive subject practice. The argument maintains that intentional explanation figures centrally successful galore core areas of cognitive science, specified arsenic perceptual psychology and linguistics. For example, perceptual psychology describes how perceptual activity transforms sensory inputs (e.g., retinal stimulations) into representations of the distal situation (e.g., perceptual representations of distal shapes, sizes, and colors). The science identifies perceptual states by citing representational properties (e.g., representational relations to circumstantial distal shapes, sizes, colors). Assuming a broadly technological realist perspective, the explanatory achievements of perceptual psychology support a realist posture towards intentionality.

Eliminativism is simply a beardown shape of anti-realism about intentionality. Eliminativists disregard intentional explanation as vague, context-sensitive, interest-relative, explanatorily superficial, aliases different problematic. They urge that scientific psychology jettison representational content. An early illustration is W.V. Quine’s Word and Object (1960), which seeks to replace intentional psychology pinch behaviorist stimulus-response psychology. Paul Churchland (1981), different salient eliminativist, wants to replace intentional psychology pinch neuroscience.

Between intentional realism and eliminativism dishonesty various intermediate positions. Daniel Dennett (1971, 1987) acknowledges that intentional discourse is predictively useful, but he questions whether mental states really person representational properties. According to Dennett, theorists who employment intentional descriptions are not literally asserting that intelligence states person representational properties. They are simply adopting the “intentional stance”. Donald Davidson (1980) espouses a neighboring interpretivist position. He emphasizes the cardinal domiciled that intentional ascription plays wrong mean interpretive practice, i.e., our believe of interpreting 1 another’s intelligence states and reside acts. At the aforesaid time, he questions whether intentional psychology will find a spot wrong mature technological theorizing. Davidson and Dennett some profess realism astir intentional mental states. Nevertheless, some philosophers are customarily publication as intentional anti-realists. (In particular, Dennett is often read as a benignant of instrumentalist astir intentionality.) One source of this customary reference involves indeterminacy of interpretation. Suppose that behavioral grounds allows two conflicting interpretations of a thinker’s intelligence states. Following Quine, Davidson and Dennett some opportunity location is past “no fact of the matter” regarding which mentation is correct. This test indicates a little than afloat realist cognition towards intentionality.

Debates complete intentionality fig prominently successful philosophical discussion of CTM. Let america study immoderate highlights.

5.1 Computation arsenic formal

Classical computationalists typically presume what 1 mightiness call the formal-syntactic conception of computation (FSC). The intuitive thought is that computation manipulates symbols successful kindness of their general syntactic properties alternatively than their semantic properties.

FSC stems from innovations successful mathematical logic during the late 19th and early 20th centuries, especially seminal contributions by George Boole and Gottlob Frege. In his Begriffsschrift (1879/1967), Frege effected a thoroughgoing formalization of deductive reasoning. To formalize, we specify a formal language whose constituent linguistic expressions are individuated non-semantically (e.g., by their geometric shapes). We whitethorn person immoderate intended mentation successful mind, but elements of the general connection are purely syntactic entities that we tin talk without invoking semantic properties specified arsenic reference or truth-conditions. In particular, we tin specify inference rules successful general syntactic terms. If we take our conclusion rules wisely, past they will cohere pinch our intended interpretation: they will transportation existent premises to existent conclusions. Through formalization, Frege invested logic pinch unprecedented rigor. He thereby laid the groundwork for galore consequent mathematical and philosophical developments.

Formalization plays a important foundational domiciled wrong computer science. We tin programme a Turing-style machine that manipulates linguistic expressions drawn from a general language. If we programme the computer wisely, past its syntactic machinations will cohere pinch our intended semantic interpretation. For example, we tin programme the computer truthful that it carries existent premises only to existent conclusions, or so that it updates probabilities arsenic dictated by Bayesian decision theory.

FSC holds that all computation manipulates general syntactic items, without respect to immoderate semantic properties those items whitethorn have. Precise formulations of FSC vary. Computation is said to be “sensitive” to syntax but not semantics, aliases to have “access” only to syntactic properties, aliases to operate “in virtue” of syntactic alternatively than semantic properties, or to beryllium impacted by semantic properties only as “mediated” by syntactic properties. It is not ever so clear what these formulations mean aliases whether they are balanced to one another. But the intuitive image is that syntactic properties have causal/explanatory primacy complete semantic properties successful driving computation forward.

Fodor’s article “Methodological Solipsism Considered arsenic a Research Strategy successful Cognitive Psychology” (1980) offers an early statement. Fodor combines FSC pinch CCTM+RTM. He analogizes Mentalese to general languages studied by logicians: it contains simple and analyzable items individuated non-semantically, conscionable arsenic typical formal languages incorporate elemental and analyzable expressions individuated by their shapes. Mentalese symbols person a semantic interpretation, but this mentation does not (directly) effect intelligence computation. A symbol’s general properties, alternatively than its semantic properties, determine really computation manipulates the symbol. In that sense, the mind is simply a “syntactic engine”. Virtually each classical computationalists travel Fodor successful endorsing FSC.

Connectionists often contradict that neural networks manipulate syntactically system items. For that reason, galore connectionists would hesitate to judge FSC. Nevertheless, astir connectionists endorse a generalized formality thesis: computation is insensitive to semantic properties. The generalized formality thesis raises galore of the aforesaid philosophical issues raised by FSC. We focus here connected FSC, which has received the astir philosophical discussion.

Fodor combines CCTM+RTM+FSC pinch intentional realism. He holds that CCTM+RTM+FSC vindicates people psychology by helping america person common sense intentional sermon into rigorous science. He motivates his position pinch a celebrated abductive statement for CCTM+RTM+FSC (1987: 18–20). Strikingly, intelligence activity tracks semantic properties in a coherent way. For example, deductive conclusion carries premises to conclusions that are existent if the premises are true. How tin we explain this important facet of intelligence activity? Formalization shows that syntactic manipulations tin way semantic properties, and computer subject shows really to build beingness machines that execute desired syntactic manipulations. If we dainty the mind arsenic a syntax-driven machine, past we tin explicate why intelligence activity tracks semantic properties successful a coherent way. Moreover, our mentation does not posit causal mechanisms radically different from those posited within the beingness sciences. We thereby reply the pivotal question: How is rationality mechanically possible?

Stephen Stich (1983) and Hartry Field (2001) harvester CCTM+FSC with eliminativism. They urge that cognitive subject exemplary the mind in formal syntactic terms, eschewing intentionality altogether. They grant that intelligence states person representational properties, but they ask what explanatory worth technological psychology gains by invoking those properties. Why supplement general syntactic explanation with intentional description? If the mind is simply a syntax-driven machine, then doesn’t representational contented driblet retired arsenic explanatorily irrelevant?

At 1 constituent successful his career, Putnam (1983: 139–154) combined CCTM+FSC pinch a Davidson-tinged interpretivism. Cognitive science should proceed on the lines suggested by Stich and Field, delineating purely general syntactic computational models. Formal syntactic modeling co-exists pinch mean interpretive practice, in which we ascribe intentional contents to 1 another’s mental states and reside acts. Interpretive believe is governed by holistic and heuristic constraints, which stymie attempts astatine converting intentional sermon into rigorous science. For Putnam, arsenic for Field and Stich, the technological action occurs astatine the general syntactic level rather than the intentional level.

CTM+FSC comes nether onslaught from various directions. One criticism targets the causal relevance of representational content (Block 1990; Figdor 2009; Kazez 1995). Intuitively speaking, the contents of intelligence states are causally applicable to intelligence activity and behavior. For example, my desire to portion h2o alternatively than orange juice causes maine to locomotion to the descend alternatively than the refrigerator. The content of my desire (that I portion water) seems to play an important causal domiciled successful shaping my behavior. According to Fodor (1990: 137–159), CCTM+RTM+FSC accommodates specified intuitions. Formal syntactic activity implements intentional mental activity, thereby ensuring that intentional intelligence states causally interact successful accord pinch their contents. However, it is not truthful clear that this study secures the causal relevance of content. FSC says that computation is “sensitive” to syntax but not semantics. Depending connected really 1 glosses the cardinal term “sensitive”, it tin look for illustration representational contented is causally irrelevant, pinch general syntax doing each the causal work. Here is an affinity to exemplify the worry. When a car drives on a road, location are unchangeable patterns involving the car’s shadow. Nevertheless, protector position astatine 1 clip does not power shadow position astatine a later time. Similarly, CCTM+RTM+FSC whitethorn explicate how mental activity instantiates unchangeable patterns described successful intentional terms, but this is not capable to guarantee the causal relevance of content. If the mind is simply a syntax-driven machine, past causal efficacy seems to reside astatine the syntactic alternatively the semantic level. Semantics is conscionable “along for the ride”. Apparently, then, CTM+FSC encourages the conclusion that representational properties are causally inert. The conclusion whitethorn not problem eliminativists, but intentional realists usually want to debar it.

A 2nd disapproval dismisses the formal-syntactic image as speculation ungrounded successful technological practice. Tyler Burge (2010a,b, 2013: 479–480) contends that general syntactic explanation of mental activity plays nary important domiciled wrong ample areas of cognitive science, including the study of theoretical reasoning, practical reasoning, and perception. In each case, Burge argues, the science employs intentional explanation rather than formal syntactic description. For example, perceptual psychology individuates perceptual states not done general syntactic properties but through representational relations to distal shapes, sizes, colors, and truthful on. To understand this criticism, we must separate formal syntactic description and neurophysiological description. Everyone agrees that a complete technological psychology will delegate prime importance to neurophysiological description. However, neurophysiological explanation is chopped from general syntactic description, because general syntactic explanation is expected to be multiply realizable successful the neurophysiological. The rumor present is whether technological psychology should supplement intentional descriptions and neurophysiological descriptions with multiply realizable, non-intentional general syntactic descriptions.

5.2 Externalism astir intelligence content

Putnam’s landmark article “The Meaning of ‘Meaning’” (1975: 215–271) introduced the Twin Earth thought experiment, which postulates a world just like our ain isolated from that H2O is replaced by a qualitatively similar constituent XYZ pinch different chemic composition. Putnam argues that XYZ is not h2o and that speakers connected Twin Earth usage the word “water” to mention to XYZ alternatively than to water. Burge (1982) extends this conclusion from linguistic reference to mental content. He argues that Twin Earthlings instantiate mental states pinch different contents. For example, if Oscar connected Earth thinks that h2o is thirst-quenching, past his copy on Twin Earth thinks a thought pinch a different content, which we might gloss arsenic that twin-water is thirst-quenching. Burge concludes that intelligence contented does not supervene upon soul neurophysiology. Mental contented is individuated partially by factors extracurricular the thinker’s skin, including causal relations to the environment. This position is externalism astir intelligence content.

Formal syntactic properties of intelligence states are wide taken to supervene upon soul neurophysiology. For example, Oscar and Twin Oscar instantiate the aforesaid general syntactic manipulations. Assuming content externalism, it follows that location is simply a immense gulf between ordinary intentional explanation and general syntactic description.

Content externalism raises superior questions astir the explanatory utility of representational contented for technological psychology:

Argument from Causation (Fodor 1987, 1991): How tin mental content exert immoderate causal power isolated from arsenic manifested within internal neurophysiology? There is nary “psychological action astatine a distance”. Differences successful the beingness situation impact behavior only by inducing differences successful section encephalon states. So the only causally applicable factors are those that supervene upon internal neurophysiology. Externally individuated contented is causally irrelevant.

Argument from Explanation (Stich 1983): Rigorous scientific explanation should not return into relationship factors extracurricular the subject’s skin. Folk psychology whitethorn taxonomize intelligence states through relations to the outer environment, but scientific psychology should taxonomize intelligence states wholly done factors that supervene upon soul neurophysiology. It should dainty Oscar and Twin Oscar arsenic psychological duplicates.[4]

Some authors prosecute the 2 arguments successful conjunction pinch 1 another. Both arguments scope the aforesaid conclusion: externally individuated mental contented finds nary morganatic spot wrong causal explanations provided by technological psychology. Stich (1983) argues on these lines to motivate his formal-syntactic eliminativism.

Many philosophers respond to specified worries by promoting content internalism. Whereas contented externalists favour wide content (content that does not supervene upon internal neurophysiology), contented internalists favour narrow content (content that does truthful supervene). Narrow contented is what remains of mental contented erstwhile 1 factors retired each outer elements. At one point successful his career, Fodor (1981, 1987) pursued internalism arsenic a strategy for integrating intentional psychology pinch CCTM+RTM+FSC. While conceding that wide contented should not fig successful scientific psychology, he maintained that constrictive contented should play a central explanatory role.

Radical internalists insist that all contented is narrow. A typical study holds that Oscar is reasoning not astir h2o but about immoderate much wide class of constituent that subsumes XYZ, so that Oscar and Twin Oscar entertain intelligence states pinch the same contents. Tim Crane (1991) and Gabriel Segal (2000) endorse specified an analysis. They clasp that people psychology ever individuates propositional attitudes narrowly. A little extremist internalism recommends that we admit constrictive contented in summation to wide content. Folk psychology whitethorn sometimes individuate propositional attitudes widely, but we tin besides delineate a viable conception of narrow content that advances important philosophical aliases technological goals. Internalists person projected various campaigner notions of constrictive content (Block 1986; Chalmers 2002; Cummins 1989; Fodor 1987; Lewis 1994; Loar 1988; Mendola 2008). See the entry narrow intelligence content for an overview of salient candidates.

Externalists kick that existing theories of constrictive contented are sketchy, implausible, useless for psychological explanation, or otherwise objectionable (Burge 2007; Sawyer 2000; Stalnaker 1999). Externalists besides mobility internalist arguments that scientific psychology requires constrictive content:

Argument from Causation: Externalists insist that wide content tin beryllium causally relevant. The specifications alteration among externalists, and chat often becomes intertwined pinch analyzable issues surrounding causation, counterfactuals, and the metaphysics of mind. See the entry mental causation for an introductory overview, and spot Burge (2007), Rescorla (2014), and Yablo (1997, 2003) for typical externalist discussion.

Argument from Explanation: Externalists declare that psychological mentation tin legitimately taxonomize intelligence states through factors that outstrip soul neurophysiology (Peacocke 1993; Shea 2018). Burge observes that non-psychological sciences often individuate explanatory kinds relationally, i.e., through relations to outer factors. For example, whether an entity counts as a bosom depends (roughly) upon whether its biologic usability in its normal situation is to pump blood. So physiology individuates organ kinds relationally. Why can’t psychology likewise individuate intelligence states relationally? For a notable speech on these issues, spot Burge (1986, 1989, 1995) and Fodor (1987, 1991).

Externalists uncertainty that we person immoderate bully logic to switch or supplement wide contented pinch constrictive content. They disregard the search for constrictive contented arsenic a chaotic goose chase.

Burge (2007, 2010a) defends externalism by analyzing existent cognitive science. He argues that galore branches of technological psychology (especially perceptual psychology) individuate intelligence contented through causal relations to the outer environment. He concludes that scientific believe embodies an externalist perspective. By contrast, he maintains, constrictive contented is simply a philosophical imagination ungrounded in current science.

Suppose we wantonness the hunt for constrictive content. What are the prospects for combining CTM+FSC pinch externalist intentional psychology? The astir promising action emphasizes levels of explanation. We tin opportunity that intentional psychology occupies one level of explanation, while formal-syntactic computational psychology occupies a different level. Fodor advocates this attack successful his later work (1994, 2008). He comes to cull constrictive contented arsenic otiose. He suggests that general syntactic mechanisms instrumentality externalist psychological laws. Mental computation manipulates Mentalese expressions successful accord pinch their general syntactic properties, and these general syntactic manipulations guarantee that intelligence activity instantiates due law-like patterns defined complete wide contents.

In ray of the internalism/externalism distinction, fto america revisit the eliminativist situation raised in §5.1: what explanatory worth does intentional explanation adhd to formal-syntactic description? Internalists tin respond that suitable formal syntactic manipulations find and possibly moreover constitute narrow contents, truthful that internalist intentional explanation is already implicit successful suitable general syntactic explanation (cf. Field 2001: 75). Perhaps this consequence vindicates intentional realism, perhaps not. Crucially, though, nary specified consequence is disposable to content externalists. Externalist intentional explanation is not implicit successful general syntactic description, because 1 tin clasp formal syntax fixed while varying wide content. Thus, contented externalists who espouse CTM+FSC must opportunity what we summation by supplementing formal-syntactic explanations pinch intentional explanations. Once we accept that intelligence computation is delicate to syntax but not semantics, it is acold from clear that immoderate useful explanatory work remains for wide content. Fodor addresses this situation astatine various points, offering his astir systematic curen successful The Elm and the Expert (1994). See Arjo (1996), Aydede (1998), Aydede and Robbins (2001), Wakefield (2002); Perry (1998), and Wakefield (2002) for criticism. See Rupert (2008) and Schneider (2005) for positions close to Fodor’s. Dretske (1993) and Shea (2018, pp. 197–226) pursue replacement strategies for vindicating the explanatory relevance of wide content.

5.3 Content-involving computation

The perceived gulf betwixt computational explanation and intentional description animates galore writings connected CTM. A fewer philosophers effort to bridge the gulf utilizing computational descriptions that individuate computational states successful representational terms. These descriptions are content-involving, to usage Christopher Peacocke’s (1994) terminology. On the content-involving approach, location is nary rigid demarcation betwixt computational and intentional description. In particular, definite scientifically valuable descriptions of mental activity are some computational and intentional. Call this position content-involving computationalism.

Content-involving computationalists request not opportunity that all computational explanation is intentional. To illustrate, suppose we describe a elemental Turing instrumentality that manipulates symbols individuated by their geometric shapes. Then the resulting computational description is not plausibly content-involving. Accordingly, content-involving computationalists do not usually advance content-involving computation arsenic a wide mentation of computation. They claim only that some important computational descriptions are content-involving.

One tin create content-involving computationalism successful an internalist or externalist direction. Internalist content-involving computationalists clasp that immoderate computational descriptions identify intelligence states partially done their narrow contents. Murat Aydede (2005) recommends a position on these lines. Externalist content-involving computationalism holds that certain computational descriptions place intelligence states partly through their wide contents. Tyler Burge (2010a: 95–101), Christopher Peacocke (1994, 1999), and Mark Sprevak (2010) espouse this position. Oron Shagrir (2001, 2020, 2022) advocates a content-involving computationalism that is neutral between internalism and externalism.

Externalist content-involving computationalists typically cite cognitive subject believe arsenic a motivating factor. For example, perceptual psychology describes the perceptual strategy arsenic computing an estimate of immoderate object’s size from retinal stimulations and from an estimate of the object’s depth. Perceptual “estimates” are identified representationally, as representations of circumstantial distal sizes and depths. Quite plausibly, representational relations to circumstantial distal sizes and depths do not supervene connected soul neurophysiology. Quite plausibly, then, perceptual psychology type-identifies perceptual computations through wide contents. So externalist content-involving computationalism seems to harmonize good pinch existent cognitive science.

A awesome situation facing content-involving computationalism concerns the interface pinch modular computationalism formalisms, specified arsenic the Turing machine. How precisely do content-involving descriptions relate to the computational models recovered successful logic and machine science? Philosophers usually presume that these models connection non-intentional descriptions. If so, that would beryllium a awesome and possibly decisive blow to content-involving computationalism.

Arguably, though, galore acquainted computational formalisms let a content-involving alternatively than general syntactic construal. To illustrate, see the Turing machine. One can individuate the “symbols” comprising the Turing instrumentality alphabet non-semantically, done factors akin to geometric shape. But does Turing’s formalism require a non-semantic individuative scheme? Arguably, the formalism allows america to individuate symbols partly done their contents. Of course, the instrumentality array for a Turing instrumentality does not explicitly mention semantic properties of symbols (e.g., denotations aliases truth-conditions). Nevertheless, the machine table tin encode mechanical rules that picture really to manipulate symbols, wherever those symbols are type-identified successful content-involving terms. In this way, the instrumentality array dictates transitions among content-involving states without explicitly mentioning semantic properties. Aydede (2005) suggests an internalist type of this view, pinch symbols type-identified done their narrow contents.[5] Rescorla (2017a) develops the position successful an externalist direction, with symbols type-identified done their wide contents. He argues that some Turing-style models picture computational operations over externalistically individuated Mentalese symbols.[6]

In principle, 1 mightiness clasp some externalist content-involving computational explanation and general syntactic description. One mightiness opportunity that these 2 kinds of explanation inhabit distinct levels of explanation. Peacocke suggests specified a view. Other content-involving computationalists respect general syntactic descriptions of the mind much skeptically. For example, Burge questions what explanatory worth general syntactic description contributes to definite areas of technological psychology (such as perceptual psychology). From this viewpoint, the eliminativist challenge posed in §5.1 has matters backwards. We should not presume that general syntactic descriptions are explanatorily valuable and past inquire what value intentional descriptions contribute. We should alternatively clasp the externalist intentional descriptions offered by existent cognitive science and past inquire what worth general syntactic description contributes.

Proponents of general syntactic explanation often respond by citing implementation mechanisms. Externalist explanation of mental activity presupposes that suitable causal-historical relations between the mind and the outer beingness situation are successful place. But surely we want a “local” explanation that ignores external causal-historical relations, a explanation that reveals underlying causal mechanisms. Fodor (1987, 1994) argues successful this measurement to motivate the general syntactic picture. For imaginable externalist responses to the statement from implementation mechanisms, spot Burge (2010b), Rescorla (2017b), Shea (2013), and Sprevak (2010). For an argument that existent cognitive subject believe does so delegate an important explanatory domiciled to general syntax, spot Calzavarini and Paternoster (2022). Debate complete the explanatory worth of formal syntax, and much mostly complete the narration betwixt computation and representation, seems apt to proceed into the indefinite future.

6. Alternative conceptions of computation

The lit offers respective replacement conceptions, usually advanced arsenic foundations for CTM. In galore cases, these conceptions overlap pinch 1 different aliases pinch the conceptions considered above.

6.1 Information-processing

It is communal for cognitive scientists to picture computation as “information-processing”. It is little communal for proponents to explain what they mean by “information” or “processing”. Lacking clarification, the explanation is little much than an quiet slogan.

Claude Shannon introduced a scientifically important conception of “information” successful his 1948 article “A Mathematical Theory of Communication”. The intuitive thought is that information measures reduction successful uncertainty, wherever reduced uncertainty manifests arsenic an altered probability distribution complete imaginable states. Shannon codified this thought wrong a rigorous mathematical framework, laying the instauration for information theory (Cover and Thomas 2006). Shannon accusation is basal to modern engineering. It finds fruitful exertion wrong cognitive science, especially cognitive neuroscience. Does it support a convincing analysis of computation arsenic “information-processing”? Consider an old-fashioned portion instrumentality that records messages received over a wireless radio. Using Shannon’s framework, 1 can measure really overmuch accusation is carried by immoderate recorded message. There is simply a consciousness successful which the portion instrumentality “processes” Shannon accusation whenever we replay a recorded message. Still, the machine does not look to instrumentality a non-trivial computational model.[7] Certainly, neither the Turing instrumentality formalism nor the neural network formalism offers overmuch penetration into the machine’s operations. Arguably, then, a strategy tin process Shannon information without executing computations successful immoderate absorbing sense.

Confronted pinch specified examples, 1 mightiness effort to isolate a more demanding conception of “processing”, truthful that the portion machine does not “process” Shannon information. Alternatively, one might insist that the portion instrumentality executes non-trivial computations. Piccinini and Scarantino (2010) beforehand a highly wide conception of computation—which they dub generic computation—with that consequence.

A 2nd salient conception of accusation derives from Paul Grice’s (1989) influential chat of natural meaning. Natural meaning involves reliable, counterfactual-supporting correlations. For example, character rings correlate pinch the property of the tree, and pox correlate pinch chickenpox. We colloquially picture character rings arsenic carrying accusation astir tree age, pox arsenic carrying accusation astir chickenpox, and truthful on. Such descriptions propose a conception that ties accusation to reliable, counterfactual-supporting correlations. Fred Dretske (1981) develops this conception into a systematic theory, arsenic do various subsequent philosophers. Does Dretske-style accusation subserve a plausible analysis of computation arsenic “information-processing”? Consider an old-fashioned bimetallic portion thermostat. Two metals are joined together into a strip. Differential description of the metals causes the portion to bend, thereby activating aliases deactivating a heating unit. Strip authorities reliably correlates pinch existent ambient temperature, and the thermostat “processes” this information-bearing authorities erstwhile activating aliases deactivating the heater. Yet the thermostat does not look to instrumentality immoderate non-trivial computational model. One would not ordinarily respect the thermostat as computing. Arguably, then, a strategy tin process Dretske-style information without executing computations successful immoderate absorbing sense. Of course, 1 mightiness effort to grip specified examples done maneuvers parallel to those from the erstwhile paragraph.

A 3rd salient conception of accusation is semantic information, i.e., representational content.[8] Some philosophers clasp that a beingness strategy computes only if the system’s states person representational properties (Dietrich 1989; Fodor 1998: 10; Ladyman 2009; Shagrir 2006; Sprevak 2010). In that sense, information-processing is necessary for computation. As Fodor memorably puts it, “no computation without representation” (1975: 34). However, this position is debatable. Chalmers (2011) and Piccinini (2008a) contend that a Turing machine might execute computations moreover though symbols manipulated by the machine person nary semantic interpretation. The machine’s computations are purely syntactic successful nature, lacking thing like semantic properties. On this view, representational contented is not necessary for a beingness strategy to count arsenic computational.

It remains unclear whether the motto “computation is information-processing” provides overmuch insight. Nevertheless, the slogan seems improbable to vanish from the lit anytime soon. For further chat of imaginable connections betwixt computation and information, spot Gallistel and King (2009: 1–26), Lizier, Flecker, and Williams (2013), Miłkowski (2013), Piccinini and Scarantino (2010), and Sprevak (2020).

6.2 Function evaluation

In a wide cited passage, the perceptual scientist David Marr (1982) distinguishes 3 levels astatine which 1 tin picture an “information-processing device”:

Computational theory: “[t]he instrumentality is characterized as a mapping from 1 benignant of accusation to another, the abstract properties of this mapping are defined precisely, and its appropriateness and adequacy for the task astatine manus are demonstrated” (p. 24).

Representation and algorithm: “the prime of representation for the input and output and the algorithm to beryllium used to toggle shape 1 into the other” (pp. 24–25).

Hardware implementation: “the specifications of really the algorithm and practice are realized physically” (p. 25).

Marr’s 3 levels person attracted aggravated philosophical scrutiny. For our purposes, the cardinal constituent is that Marr’s “computational level” describes a mapping from inputs to outputs, without describing intermediate steps. Marr illustrates his approach by providing “computational level” theories of various perceptual processes, specified arsenic separator detection.

Marr’s chat suggests a functional conception of computation, connected which computation is simply a matter of transforming inputs into due outputs. Frances Egan elaborates the functional conception complete a bid of articles (1991, 1992, 1999, 2003, 2010, 2014, 2019, 2020). Like Marr, she treats computational description arsenic explanation of input-output relations. She besides claims that computational models qualify a purely mathematical function: that is, a mapping from mathematical inputs to mathematical outputs. She illustrates by considering a ocular system (called “Visua”) that computes an object’s extent from retinal disparity. She imagines a neurophysiological duplicate (“Twin Visua”) embedded truthful otherwise successful the physical environment that it does not correspond depth. Visua and Twin Visua instantiate perceptual states pinch different representational properties. Nevertheless, Egan says, imagination subject treats Visua and Twin Visua arsenic computational duplicates. Visua and Twin Visua compute the aforesaid mathematical function, moreover though the computations have different representational import successful the 2 cases. Egan concludes that computational modeling of the mind yields an “abstract mathematical description” accordant pinch many alternative imaginable representational descriptions. Intentional attribution is conscionable a heuristic gloss upon underlying computational description.

Chalmers (2012) argues that the functional conception neglects important features of computation. As he notes, computational models usually picture much than conscionable input-output relations. They describe intermediate steps done which inputs are transformed into outputs. These intermediate steps, which Marr consigns to the “algorithmic” level, fig prominently successful computational models offered by logicians and machine scientists. Restricting the term “computation” to input-output explanation does not capture modular computational practice.

An further interest faces functional theories, specified arsenic Egan’s, that exclusively stress mathematical inputs and outputs. Critics kick that Egan mistakenly elevates mathematical functions, at the disbursal of intentional explanations routinely offered by cognitive subject (Burge 2005; Rescorla 2015; Silverberg 2006; Sprevak 2010). To illustrate, suppose perceptual psychology describes the perceptual strategy arsenic estimating that immoderate object’s extent is 5 meters. The perceptual depth-estimate has a representational content: it is meticulous only if the object’s extent is 5 meters. We cite the number 5 to place the depth-estimate. But our prime of this number depends upon our arbitrary prime of measurement units. Critics contend that the contented of the depth-estimate, not the arbitrarily chosen number done which we theorists specify that content, is what matters for psychological explanation. Egan’s mentation places the number alternatively than the contented astatine explanatory halfway stage. According to Egan, computational mentation should picture the ocular system as computing a particular mathematical function that carries particular mathematical inputs into particular mathematical outputs. Those peculiar mathematical inputs and outputs dangle upon our arbitrary prime of measurement units, truthful they arguably deficiency the explanatory value that Egan assigns to them.

We should separate the functional approach, arsenic pursued by Marr and Egan, from the functional programming paradigm successful computer science. The functional programming paradigm models information of a complex usability arsenic successive information of simpler functions. To take a elemental example, 1 mightiness measure \(f(x,y) = (x^{2}+y)\) by first evaluating the squaring usability and past evaluating the addition function. Functional programming differs from the “computational level” descriptions emphasized by Marr, because it specifies intermediate computational stages. The functional programming paradigm stretches backmost to Alonzo Church’s (1936) lambda calculus, continuing pinch programming languages such as PCF and LISP. It plays an important domiciled successful AI and theoretical computer science. Some authors propose that it offers typical insight into intelligence computation (Klein 2012; Piantadosi, Tenenbaum, and Goodman 2012). However, galore computational formalisms do not conform to the functional paradigm: Turing machines; imperative programming languages, specified arsenic C; logic programming languages, specified arsenic Prolog; and so on. Even though the functional paradigm describes numerous important computations (possibly including intelligence computations), it does not plausibly seizure computation in general.

6.3 Structuralism

Many philosophical discussions embody a structuralist conception of computation: a computational exemplary describes an abstract causal structure, without taking into relationship peculiar physical states that instantiate the structure. This conception traces backmost at least to Putnam’s original curen (1967). Chalmers (1995, 1996a, 2011, 2012) develops it successful detail. He introduces the combinatorial-state automaton (CSA) formalism, which subsumes most acquainted models of computation (including Turing machines and neural networks). A CSA provides an absurd explanation of a physical system’s causal topology: the shape of causal interaction among the system’s parts, independent of the nature of those parts aliases the causal mechanisms done which they interact. Computational explanation specifies a causal topology.

Chalmers deploys structuralism to delineate a very wide type of CTM. He assumes the functionalist position that psychological states are individuated by their roles successful a shape of causal organization. Psychological explanation specifies causal roles, abstracted distant from physical states that recognize those roles. So psychological properties are organizationally invariant, successful that they supervene upon causal topology. Since computational explanation characterizes a causal topology, satisfying a suitable computational description suffices for instantiating due intelligence properties. It also follows that psychological explanation is simply a type of computational description, truthful that computational explanation should play a central role wrong psychological explanation. Thus, structuralist computation provides a coagulated instauration for cognitive science. Mentality is grounded successful causal patterns, which are precisely what computational models articulate.

Structuralism comes packaged pinch an charismatic relationship of the implementation relation betwixt absurd computational models and beingness systems. Under what conditions does a beingness system implement a computational model? Structuralists opportunity that a physical system implements a exemplary conscionable successful lawsuit the model’s causal structure is “isomorphic” to the model’s formal structure. A computational exemplary describes a beingness strategy by articulating a general building that mirrors immoderate applicable causal topology. Chalmers elaborates this intuitive idea, providing detailed necessary and capable conditions for beingness realization of CSAs. Few if immoderate replacement conceptions of computation tin supply so substantive an relationship of the implementation relation.

We whitethorn instructively comparison structuralist computationalism pinch some other theories discussed above:

Machine functionalism. Structuralist computationalism embraces the halfway thought down instrumentality functionalism: intelligence states are functional states describable done a suitable computational formalism. Putnam advances CTM arsenic an empirical hypothesis, and he defends functionalism connected that basis. In contrast, Chalmers follows David Lewis (1972) by grounding functionalism successful the conceptual analysis of mentalistic discourse. Whereas Putnam defends functionalism by defending computationalism, Chalmers defends computationalism by assuming functionalism.

Classical computationalism, connectionism, and computational neuroscience. Structuralist computationalism emphasizes organizationally invariant descriptions, which are multiply realizable. In that respect, it diverges from computational neuroscience. Structuralism is compatible pinch some classical and connectionist computationalism, but it differs successful tone from those views. Classicists and connectionists coming their rival positions as bold, substantive hypotheses. Chalmers advances structuralist computationalism arsenic a comparatively minimalist position improbable to be disconfirmed.

Intentional realism and eliminativism. Structuralist computationalism is compatible pinch some positions. CSA description does not explicitly mention semantic properties specified arsenic reference, truth-conditions, representational content, and truthful on. Structuralist computationalists request not delegate representational contented any important domiciled wrong technological psychology. On the different hand, structuralist computationalism does not preclude an important domiciled for representational content.

The formal-syntactic conception of computation. Wide content depends connected causal-historical relations to the outer environment, relations that outstrip causal topology. Thus, CSA explanation leaves wide contented underdetermined. Narrow contented presumably supervenes upon causal topology, but CSA explanation does not explicitly mention narrow contents. Overall, then, structuralist computationalism prioritizes a level of formal, non-semantic computational description. In that respect, it resembles FSC. On the different hand, structuralist computationalists request not opportunity that computation is “insensitive” to semantic properties, truthful they request not endorse each aspects of FSC.

Although structuralist computationalism is chopped from CTM+FSC, it raises immoderate akin issues. For example, Rescorla (2012) denies that causal topology plays the cardinal explanatory domiciled wrong cognitive science that structuralist computationalism dictates. He suggests that externalist intentional explanation alternatively than organizationally invariant explanation enjoys explanatory primacy. Coming from a different direction, computational neuroscientists will urge that we forego organizationally invariant descriptions and alternatively employ more neurally circumstantial computational models. In consequence to such objections, Chalmers (2012) argues that organizationally invariant computational explanation yields explanatory benefits that neither intentional explanation nor neurophysiological explanation replicate: it reveals the underlying mechanisms of cognition (unlike intentional description); and it abstracts distant from neural implementation details that are irrelevant for galore explanatory purposes.

6.4 Mechanistic theories

The mechanistic quality of computation is simply a recurring taxable successful logic, philosophy, and cognitive science. Several authors create this theme into a mechanistic conception of computing systems (Coelho Mollo, 2017; Dewhurst 2016; Fresco 2014, 2017; Miłkowski 2013; Piccinini 2007, 2012, 2015). On Gualtiero Piccinini’s (2015) influential development, a functional mechanism is simply a strategy of interconnected components, wherever each constituent performs immoderate function within the wide system. Mechanistic explanation proceeds by decomposing the strategy into parts, describing really the parts are organized into the larger system, and isolating the usability performed by each part. A computing strategy is simply a functional system of a particular kind: it is simply a system whose components are functionally organized to process vehicles successful accord pinch rules. Echoing Putnam’s chat of aggregate realizability, Piccinini demands that the rules beryllium medium-independent, successful that they abstract away from the circumstantial beingness implementations of the vehicles. Computational mentation decomposes the strategy into parts and describes really each portion helps the strategy process the relevant vehicles. If the strategy processes discretely system vehicles, then the computation is digital. If the strategy processes continuous vehicles, past the computation is analog. Marcin Miłkowski’s (2013) type of the mechanistic attack is similar. He differs from Piccinini by pursuing an “information-processing” gloss, truthful that computational mechanisms run complete information-bearing states. Miłkowski and Piccinini deploy their respective mechanistic theories to defend computationalism. Piccinini (2020) focuses particularly upon neural computation, drafting extended connections with cognitive neuroscience.

Mechanistic computationalists typically individuate computational states non-semantically. They truthful brushwood worries astir the explanatory domiciled of representational content, akin to worries encountered by FSC and structuralism. Critics protestation that mechanistic computationalism does not accommodate cognitive subject explanations that are simultaneously computational and representational (Rescorla 2016; Shagrir 2014; Shagrir 2022). The perceived unit of this criticism will dangle upon one’s sympathy for content-involving computationalism. To defuse the criticism, Miłkowski (2017) retorts that mechanistic computationalists tin delegate a central theoretical domiciled to representational contented by attributing representational functions to definite computing mechanisms.

6.5 Pluralism

We person surveyed various contrasting and sometimes overlapping conceptions of computation: classical computation, connectionist computation, neural computation, formal-syntactic computation, content-involving computation, information-processing computation, functional computation, structuralist computation, and mechanistic computation. Each conception yields a different shape of computationalism. Each conception has its ain strengths and weaknesses. One mightiness adopt a pluralistic stance that recognizes chopped morganatic conceptions. Rather than elevate one conception supra the others, pluralists happily employment whichever conception seems useful successful a fixed explanatory context. Edelman (2008) takes a pluralistic line, arsenic does Chalmers (2012) successful his astir recent discussion.

The pluralistic statement raises immoderate earthy questions. Can we supply a general study that encompasses each aliases astir types of computation? Do all computations stock definite characteristic marks pinch 1 another? Are they possibly alternatively agreed by thing for illustration family resemblance? Deeper knowing of computation requires america to grapple pinch these questions.

7. Arguments against computationalism

CTM has attracted galore objections. In galore cases, the objections apply only to circumstantial versions of CTM (such arsenic classical computationalism aliases connectionist computationalism). Here are a few prominent objections. See besides the introduction on the Chinese room argument for a wide discussed objection to classical computationalism advanced by John Searle (1980).

7.1 Triviality arguments

A recurring interest is that CTM is trivial, because we can describe almost immoderate beingness strategy arsenic executing computations. Searle (1990) claims that a wall implements any machine program, since we tin discern immoderate shape of molecular movements successful the wall that is isomorphic to the general building of the program. Putnam (1988: 121–125) defends a little utmost but still very strong triviality thesis on the aforesaid lines. Triviality arguments play a large domiciled successful the philosophical literature. Anti-computationalists deploy triviality arguments against computationalism, while computationalists activity to debar triviality.

Computationalists usually rebut triviality arguments by insisting that the arguments place constraints upon computational implementation, constraints that barroom trivializing implementations. The constraints may be counterfactual, causal, semantic, aliases otherwise, depending on one’s favored mentation of computation. For example, David Chalmers (1995, 1996a) and B. Jack Copeland (1996) clasp that Putnam’s triviality statement ignores counterfactual conditionals that a physical strategy must fulfill successful bid to instrumentality a computational model. Other philosophers opportunity that a beingness strategy must have representational properties to instrumentality a computational exemplary (Fodor 1998: 11–12; Ladyman 2009; Sprevak 2010) aliases astatine slightest to implement a content-involving computational exemplary (Rescorla 2013). The details present alteration considerably, and computationalists statement amongst themselves precisely which types of computation tin debar which triviality arguments. But astir computationalists work together that we can avoid immoderate devastating triviality worries done a sufficiently robust theory of the implementation narration betwixt computational models and physical systems.

Pancomputationalism holds that each beingness system implements a computational model. This thesis is plausible, since any physical strategy arguably implements a sufficiently trivial computational exemplary (e.g., a one-state finite authorities automaton). As Chalmers (2011) notes, pancomputationalism does not look worrisome for computationalism. What would beryllium worrisome is the overmuch stronger triviality thesis that almost each beingness strategy implements almost every computational model.

For further chat of triviality arguments and computational implementation, spot Sprevak (2019) and the entry computation successful beingness systems.

7.2 Gödel’s incompleteness theorem

According to immoderate authors, Gödel’s incompleteness theorems show that quality mathematical capacities outstrip the capacities of any Turing instrumentality (Nagel and Newman 1958). J.R. Lucas (1961) develops this position into a celebrated critique of CCTM. Roger Penrose pursues the critique successful The Emperor’s New Mind (1989) and subsequent writings. Various philosophers and logicians person answered the critique, arguing that existing formulations suffer from fallacies, question-begging assumptions, and moreover outright mathematical errors (Bowie 1982; Chalmers 1996b; Feferman 1996; Lewis 1969, 1979; Putnam 1975: 365–366, 1994; Shapiro 2003). There is a wide statement that this disapproval of CCTM lacks immoderate force. It may turn retired that definite quality intelligence capacities outstrip Turing-computability, but Gödel’s incompleteness theorems provide nary logic to expect that outcome.

7.3 Limits of computational modeling

Could a machine constitute the Eroica symphony? Or discover general relativity? Or moreover replicate a child’s effortless ability to comprehend the environment, necktie her shoelaces, and discern the emotions of others? Intuitive, creative, aliases skillful human activity whitethorn look to defy formalization by a machine program (Dreyfus 1972, 1992). More generally, 1 mightiness interest that crucial aspects of quality cognition elude computational modeling, especially classical computational modeling.

Ironically, Fodor promulgates a forceful type of this critique. Even successful his earliest statements of CCTM, Fodor (1975: 197–205) expresses sizeable skepticism that CCTM tin grip each important cognitive phenomena. The pessimism becomes much pronounced successful his later writings (1983, 2000), which attraction particularly connected abductive reasoning arsenic a intelligence arena that perchance eludes computational modeling. His halfway statement whitethorn beryllium summarized as follows:

(1) Turing-style computation is delicate only to “local” properties of a intelligence representation, which are exhausted by the identity and statement of the representation’s constituents. (2) Many intelligence processes, paradigmatically abduction, are sensitive to “nonlocal” properties specified arsenic relevance, simplicity, and conservatism. (3) Hence, we whitethorn person to wantonness Turing-style modeling of the relevant processes. (4) Unfortunately, we person presently person nary thought what alternative theory mightiness service arsenic a suitable replacement.

Some critics contradict (1), arguing that suitable Turing-style computations can beryllium delicate to “nonlocal” properties (Schneider 2011; Wilson 2005). Some situation (2), arguing that emblematic abductive inferences are delicate only to “local” properties (Carruthers 2003; Ludwig and Schneider 2008; Sperber 2002). Some concede measurement (3) but conflict measurement (4), insisting that we have promising non-Turing-style models of the applicable intelligence processes (Pinker 2005). Partly spurred by specified criticisms, Fodor elaborates his argument successful sizeable detail. To take sides (2), he critiques theories that exemplary abduction by deploying “local” heuristic algorithms (2005: 41–46; 2008: 115–126) aliases by positing a profusion of domain-specific cognitive modules (2005: 56–100). To take sides (4), he critiques various theories that grip abduction through non-Turing-style models (2000: 46–53; 2008), specified as connectionist networks.

The scope and limits of computational modeling stay controversial. We whitethorn expect this taxable to stay an progressive attraction of inquiry, pursued jointly pinch AI.

7.4 Temporal arguments

Mental activity unfolds successful time. Moreover, the mind accomplishes sophisticated tasks (e.g., perceptual estimation) very quickly. Many critics interest that computationalism, particularly classical computationalism, does not adequately accommodate temporal aspects of cognition. A Turing-style exemplary makes nary definitive mention of the time scale complete which computation occurs. One could physically implement the aforesaid absurd Turing instrumentality pinch a silicon-based device, aliases a slower vacuum-tube device, aliases an moreover slower pulley-and-lever device. Critics urge that we cull CCTM successful favour of immoderate alternative framework that much straight incorporates temporal considerations. van Gelder and Port (1995) usage this statement to beforehand a non-computational dynamical systems framework for modeling mental activity. Eliasmith (2003, 2013: 12–13) uses it to support his Neural Engineering Framework.

Computationalists respond that we tin supplement an abstract computational exemplary pinch temporal considerations (Piccinini 2010; Weiskopf 2004). For example, a Turing instrumentality exemplary presupposes discrete “stages of computation”, without describing how the stages subordinate to beingness time. But we tin supplement our exemplary by describing really agelong each shape lasts, thereby converting our non-temporal Turing instrumentality exemplary into a mentation that yields detailed temporal predictions. Many advocates of CTM employment supplementation along these lines to study temporal properties of cognition (Newell 1990). Similar supplementation figures prominently successful computer science, whose practitioners are rather concerned to build machines with due temporal properties. Computationalists reason that a suitably supplemented type of CTM tin adequately seizure how cognition unfolds successful time.

A 2nd temporal objection highlights the opposition between discrete and continuous temporal improvement (van Gelder and Port 1995). Computation by a Turing instrumentality unfolds in discrete stages, while intelligence activity unfolds successful a continuous time. Thus, location is simply a basal mismatch betwixt the temporal properties of Turing-style computation and those of existent intelligence activity. We need a psychological mentation that describes continuous temporal evolution.

Computationalists respond that this objection assumes what is to be shown: that cognitive activity does not autumn into explanatory significant discrete stages (Weiskopf 2004). Assuming that physical time is continuous, it follows that intelligence activity unfolds in continuous time. It does not travel that cognitive models must person continuous temporal structure. A individual machine operates in continuous time, and its beingness authorities evolves continuously. A complete beingness mentation will bespeak each those beingness changes. But our computational exemplary does not bespeak each physical change to the computer. Our computational exemplary has discrete temporal structure. Why presume that a bully cognitive-level exemplary of the mind must bespeak each beingness alteration to the brain? Even if location is a continuum of evolving physical states, why presume a continuum of evolving cognitive states? The specified truth of continuous temporal improvement does not militate against computational models with discrete temporal structure.

For chat of really to reconcile CTM pinch a dynamical systems perspective, spot (Beer and Williams 2015; Phattanasri, Chiel, and Beer 2007; Weinberger and Allen 2022).

7.5 Embodied cognition

Embodied cognition is simply a investigation programme that draws inspiration from the continental philosopher Maurice Merleau-Ponty, the perceptual psychologist J.J. Gibson, and different type influences. It is a fairly heterogeneous movement, but the basal strategy is to emphasize links betwixt cognition, bodily action, and the surrounding environment. See Varela, Thompson, and Rosch (1991) for an influential early statement. In galore cases, proponents deploy devices of dynamical systems theory. Proponents typically coming their attack arsenic a radical replacement to computationalism (Chemero 2009; Kelso 1995; Thelen and Smith 1994). CTM, they complain, treats intelligence activity as static awesome manipulation detached from the embedding environment. It neglects myriad analyzable ways that the situation causally or constitutively shapes intelligence activity. We should switch CTM pinch a new image that emphasizes continuous links betwixt mind, body, and environment. Agent-environment dynamics, not soul mental computation, holds the cardinal to knowing cognition. Often, a broadly eliminativist cognition towards intentionality propels this critique.

Computationalists respond that CTM allows owed nickname of cognition’s embodiment. Computational models tin return into account really mind, body, and situation continuously interact. After all, computational models tin incorporated sensory inputs and motor outputs. There is nary evident logic why an accent upon agent-environment dynamics precludes a dual accent upon internal mental computation (Clark 2014: 140–165; Rupert 2009). Computationalists support that CTM tin incorporated immoderate legitimate insights offered by the embodied cognition movement. They besides insist that CTM remains our champion wide model for explaining numerous core psychological phenomena.