Cog, an abnormal Alphabet-CSE

Did Cog afford you wise options?

How do human minds work?: The Cognitive Revolution and Paradigm Change in Cognitive Science

During the first half of the 20th century, empiricism permeated most fields related to the study of human minds, particularly epistemology and the social sciences. The pendulum swung toward empiricism at the end of the 19th century in reaction to the introspective and speculative methods that had become the standard in disciplines like psychology, psychophysics and philosophy. Based on technical advances mostly achieved in Russia and the United States, behaviorism took form, threatening to absorb philosophy of language and linguistics (e.g., respectively, Quine 1960, and Skinner 1948, 1957).  In reaction to that movement, Cognitive Science emerged as an alternative for those discontent with the reigning versions of empiricism, that is, as a rationalist alternative.

When Chomsky (1959) pounced upon Skinner's Verbal Behavior, he later reasserted his victory as a vindication of rationalism in the face of “a futile tendency in modern speculation”, stating that he did not "see any way in which his proposals can be substantially improved within the general framework of behaviorist or neobehaviorist, or, more generally, empiricist ideas that has dominated much of modern linguistics, psychology, and philosophy" (Chomsky 1967).  Noam Chomsky’s assault, backed by the research program offered alongside it (Chomsky 1957), would be followed by twenty-five years of almost completely uncontested rationalist consensus.  Thus, the Cognitive Revolution is best understood as a rationalist revolution.

Researchers in the newly delineated interdisciplinary field coincided in arguing that the mind employs syntactic processes on amodal (i.e., context-independent) structured symbol, some of which must be innate.  The computer metaphor guided the formulation of models, whereby mind is to nervous system what software is to hardware.  Conceived as a new scientific epistemology, Cognitive Science built bridges across separate disciplines.  
Though each field has its own terminology dissimilar to the others, potentially straining effective communication, academics could converge on the view that thought, reasoning, decision-making, and problem-solving are logical, syntactic, serial processes over structured symbols.  As such, it may be suggested that the rationalist framework greatly facilitated the gestation and institutional validation of Cognitive Science as a academic domain in its own right.  Human cognition could be though of as Turing Machines (Turing 1936), perhaps similar to a von Neumann architecture (von Neumann 1945), that obey George Boole's (1854) Laws of Thought, and this computational foundation worked equally well for generative linguists, cognitive psychologists, neuroscientists, computer programmers focused on artificial intelligence, and analytic philosophers fixated on the propositional calculus of inference and human reason.  Consequently, most textbook on cognition contain a few diagrams like the one below.

Models that abide by the aforementioned rationalist premises are known as classicalist or as having a Classical Cognitive Architecture (Fodor and Pylyshyn 1988). It wasn’t until the mid-80s, with the resurgence of modeling via artificial neural networks, that the rationalist hegemony began to crack at the edges, as increasing emphasis was placed on learning algorithms based on association, induction, and statistical mechanisms that for the most part attempted to do away with innate representations altogether.  This resurgence threw Cognitive Science into what Bechtel, Abrahamsen & Graham (1998) called an identity crisis, which they date from 1985 until the time of that publication.  Almost two decades later, the identity crisis remains unresolved, as this new approach has been met with fierce resistance, displaying the unnerving, painstakingly slow characteristics of a Kuhnian paradigm shift (Kuhn 1962).

In Hume Variations (2003), Jerry Fodor, the most prominent and radical rationalist philosopher of Cognitive Science alive today, rescued the Cartesian in Hume along with his naïve Faculty Psychology at the cost of sacrificing his associationist view of learning.  And of course Fodor did this since that maneuver would render Hume a rationalist and also Cartesian linguistics and reason are central to the inaugural program of Cognitive Science, a framework that Fodor helped construct from the very beginning.  Chomsky's (1966) Cartesian Linguistics traces many of the developments of his own linguistic theory, including the key distinction between surface structure and deep structure, to the Port-Royal Grammar published by Arnauld and Lancelot in 1660.  The Port-Royal Grammar and the Port-Royal Logic (Arnauld and Nicole 1662) were both heavily influenced by the work of René Descartes.  However, the evidence is quickly mounting in a way that suggests that the maneuver needed is the opposite of Fodor's, that is, to rescue the associationist theory of learning while discarding the Cartesian aspects and the folk Faculty Psychology present in Hume's philosophy of mind.

A brief comparison between the prototypical rationalist and empiricist stances is provided in the following table.

Of these positions, the rationalist / empiricist distinction in philosophy of mind rests squarely on the issue of representational nativism. The other facets (listed in mind, processes, and representations above) seem to follow from what would be needed, wanted or expected of a cognitive architecture if there were either some or no innate ideas.

That there are no innate ideas is the core tenet of empiricist philosophy of mind. Hume believed that the mind was made up of faculties, a modular association of distinct associative engines, but he left open the question of whether the faculties arise out of experience (or ‘custom’) or are innately specified (and to what extent). There are two main reasons that suggest the former option to be the case.  First, uncommitted neural networks approximate functions, both of the body and of the world, paving the way for functional organization through processes of neural auto-organization. Second, committed neural networks bootstrap one another towards the approximation of more complicated functions; as this occurs, the domain-general processes of neurons give way to domain-specific functional organizations. However, though the representations that constitute these domain-specific processes can become increasingly applicable to variable contexts, these do not become wholly amodal, that is, context-independent, because domain-specific functions are anchored in domain-general associative processes that are inherently context-dependent or modal. (See How You Know What You Know for a review of scientific research that supports the two aforementioned reasons.)

Having said this, it must be noted that neither rationalism nor empiricism actually constitute a theory of anything at all; their core is only one hypothesis – either there are some innate ideas or there are none. There is the third possibility, however, that ideas do not exist, at least not in minds, making the rationalist/empiricist debate obsolete (cf., Brooks 1991). This third option notwithstanding, even though neither empiricism nor rationalism is actually a theory of mind, it is possible to build one in the spirit of their corresponding proposition. That is what Locke, Berkeley and Hume did; it is also what Noam Chomsky did, and what Lawrence Barsalou is doing now (whose research program is stated in Barsalou 1999).

Be that as it may, the rationalist consensus that dominated Cognitive Science's first thirty years cannot be explained by mere technological or technical factors. While someone could argue that connectionism did not appear until the mid-80s because neural networks could not be artificially implemented, this claim would be historically unfounded. Bechtel, Abrahamsen & Graham (1998) pinpoint September 11, 1956 as the date of birth of Cognitive Science. Though one may be reluctant to accept such a specific date, it is clear that the inter-disciplinary field emerged around then, plus or minus a few years. However, already in 1943, McCulloch and Pitts proposed an abstract model of neurons and showed how any logical function could be represented in networks of these simple units of computation. By 1956, several research teams had tried their hand at implementing neural networks on digital computers (see, e.g., the project of Rochester, Holland, Haibt & Duda 1956 at IBM).  By the early 60's, not only had the idea been explored, Rosenblatt (1962) had even tried building artificial neural networks as actual machines, using photovoltaic cells, instead of just simulating these on digital computers.

When Cognitive Science emerged, the technological tools existed so that research could have gone the rationalist’s or the empiricist’s way, or at least remained neutral on the matter; however, as the Cognitive Revolution is best understood as a rationalist revolution, nativism was hailed, construction began on a Universal Grammar (a project that failed miserably, by the way), decision-making processes were construed as syntactic manipulations on explicit symbol structures (Newell, Shaw, and Simon 1959, Anderson 1982), and neural networks were taken as simple instruments of pattern recognition that could serve to augment a classical cognitive architecture or, at most, to implement what would ultimately be a rationalist story. Fodor & Pylyshyn (1988) were surprisingly blunt on this last point by stating that the issue of connectionism constituting a model of cognition “is a matter that was substantially put to rest about thirty years ago” when the Cognitive Revolution took place. It took thirty years of work for frustration to set in with rationalist approaches; only then would connectionism reappear, augmented by the tools of dynamical systems theory, as a viable alternative to the rationalist or classicalist conception of cognition.

Paradigm Change in Artificial Intelligence

The term ‘connectionist’ was introduced by Donald Hebb (1949) and revived by Feldman (1981) to refer to a class of neural networks that compute through the connection weights. Thousands of connectionist nets, similar to some degree or other to the schematic below, have been created since the 1950s. The wide variety of artificial neural networks is due not only to the function each has been created (and raised) to carry out, which constrains the type of inputs and outputs to which the system has access, but also to their specific architecture—the number of neuron each layer contains, the kind of connections these exhibit, the number of layers, and the class of learning algorithm that calibrate its connection weights.

A clear and very simple example of a connectionist net (seen below) was developed by McClelland and Rumelhart (1981) for word recognition. The 3-layer network proceeded from the visual features of letters to the recognition of words through localist representations of letters in the hidden layer (for a richer discussion, see McClelland 1989). Given its function and the use of localist representations, both the mode of presentation of the input and the mode of generation of the output was constrained by the features of written language, which in turn delineated the network’s design.

Borrowed from the Empirical Philosophy of Science Project at the Natural Computation Lab of the University of California, San Diego, the graph below evidences the transition from the classicalist paradigm to the connectionist by presenting the frequency of appearance (by year) of the lexical items ‘expert system’ and ‘neural network’ in peer-reviewed academic journals of Cognitive Science. It can be clearly seen that the interest in neural networks supplanted the 1980's craze for expert systems.

For those lacking knowledge on the matter, an expert system is a decision-making program that is supposed to mimic the inferences of an expert in a given field; basically, the shell of the program is an inference engine that works logically and syntactically, and this engine must be given a knowledge base, a finite set of "If X, then Y" rules the sum of which ought to allow it to perform its target function correctly most of the time.  Typically, an expert system asks you either questions or to input specific data, and using those inputs, the inference engine goes through its knowledge base to provide you an answer.  Expert systems may be created for purposes of prediction, planning, monitoring, debugging, and perhaps most prominently for diagnosis, among several other possible purposes.  WebMD's symptom checker, which you may have used once or twice, is perhaps the most well-known example; you click on what symptoms you have, its inference engine passes your data through its knowledge base, and it provides you with a list of all the sicknesses you may be suffering from.  If you have used that symptom checker more than twice in your life, you probably know how inaccurate it tends to be, even to the point of being ludicrous at times.  In stark contrast, many artificial neural networks have been created for detecting all sorts of cancers and can do with 99% accuracy, that is, better than almost any doctor, like this one for breast cancer created by a girl during her junior year of high school.  This is just one out of countless domains where empiricist approaches vastly outperform their rationalist counterparts.

As a funny digression, I once had to make an expert system for a graduate class and built a program that would ask you 16 socioeconomic and political questions, from which it would diagnose your preferred political philosophy  (e.g., anarchism, liberalism, republicanism, communism, constitutional monarchist, fascism, and so on).  My artificial intelligence professor took it with him to the School of Engineering to test it out on his students, and when I saw him again, he commented that he was impressed by how accurate it was.  It was definitely more accurate than WebMD but, then again, medical diagnosis is a way more complicated knowledge domain that contains many more possible outputs so that is an unfair comparison.  On an unrelated but also funny note, my other artificial intelligence professor told the story of how he had lost faith in artificial neural networks while at grad school when he created a system that would either approve or reject a bank loan application. He would input the demographic and personal income data as well as the loan information, and the network would respond a simple Approve or Reject.  But he created the network with a twist; he deliberately trained it with a racist data set in such a way that the network wouldn't give out any prime loans to anyone that wasn't white.  He wanted to see if the network would ever learn the error of his ways or at least acknowledge its racism, but it never did, and he said that at that moment he lost all faith in connectionist networks.  When he finished telling the story, I immediately raised my hand and said—"You do realize that that is exactly what happens with many bankers in real life, right?  You network didn't fail; it behaved like a human would."

Reframing Cognitive Science

The seeds of empiricism have been sprouting almost everywhere. The last thirty years have seen an ever-increasing portion of scientific research dedicated, even if reluctantly, to proving some of the central tenets of empiricist theory of mind or attempting to articulate mechanisms to augment it.

In artificial intelligence, connectionist architecture emerged in the 80's as a clear and feasible alternative to symbolic approaches (a.k.a., good old-fashioned artificial intelligence or GOFAI; Haugeland 1985, Dreyfus 1992). The tools of dynamical systems theory, widely used in the field of physics, bolstered connectionism to provide for a robust account of a system’s ontogenetic evolution through time (van Gelder 1999). Connectionism provided that which behaviorist lacked, powerful learning mechanisms that could account for not only how intelligent agents derive knowledge from experience but also how we can surpass that limited amount of information to conceive an unlimited amount of possibilities; furthermore, the tools of dynamical systems theory opened the possibility of seeing what goes on inside the ‘black box’, while also helping psychology get in sync with physics and neurology. In this sense, connectionism ought not to be confused with behaviorism because neural network architectures permit an agent to surpass the limited stimulus-response patterns that it encounters (Lewis and Elman 2001, Elman 1998). It should be noted, however, that connectionist computation is not synonymous with empiricism, that it is, in fact, entirely compatible with rationalist postulates, as exemplified by Optimality Theory (Prince & Smolensky 1997), an attempt to implement universal grammar via a connectionist architecture; nevertheless, this compatibility is a token truism that goes both ways and is due to the fact that artificial neural networks and Turing machines exhibit equivalent computational power inasmuch as either can implement any definable function, which is why most people simulate neural networks using common personal computers (currently, the best open-source, free software for creating your own neural network with relative ease is Emergent, a program hosted by the University of Colorado that runs on Windows, Macintosh OS's, and Linux-Ubuntu, and can be downloaded here). Looking beyond this universal computational compatibility, connectionism clearly opens the door to empiricism, and the vast majority of connectionist models do away with rationalist tenets and clearly partake of the long-standing empiricist tradition even if many of their authors aren't willing to admit this publicly because of the entrenched stigma branded into that philosophical label.

In linguistics, a clear alternative to generativism surfaced during the 1980s in the form of Cognitive Linguistics (Langacker 1987, Lakoff 1987). Though cognitive linguistics is not wholeheartedly committed to an empiricist theory of mind, its rejection of the fundamental tenets of generativism is in itself a retreat from the rationalist consensus that stood almost uncontested. Specifically, its rejection of an autonomous, modular universal grammar and its grounding of linguistic abilities in domain-general learning and associative mechanisms represent a big leap towards empiricism. Moreover, as linguistics increasingly meshes with psychology and connectionism, slowly but surely an associationist flavor that had long been wiped out by Chomsky and his followers returns to the field. In consequence, much work in linguistics is being fruitfully redirected from devising categorical acquisition schemes toward testing statistical learning algorithms for the acquisition of syntax as well as for syntax's prehistoric origins (e.g., Hazlehurst and Hutchins 1998, Hutchins and Hazlehurst 1995) and also for how grammar changes throughout history (see, e.g., Hare and Elman 1995).

In psychology, many connectionist-friendly accounts have been offered. Perhaps the most ambitious is Barsalou’s (1999) perceptual symbol systems, an account that takes a firm empiricist stance in the face of rationalist psychology by dissolving the distinction between perception and conception. Moreover, the perceptual symbol systems approach has been recently applied, though not without difficulties, to theory of discourse (Zwaan 2004) and to theory of concepts (Prinz 2002). Still, this is not the only empiricist current in psychology, as the domain of psycholinguistics has been propelled mostly by psychologists, like Elizabeth Bates and Brian MacWhinney, and has led to findings and models that are very compatible with the tenet of empiricism (see, e.g., Thelen and Bates 2003, Tomasello 2006, Goldberg 2004, MacWhinney 2013).  Not to mention that many of the early proponents of the parallel distributed processing (or PDP) approach to Cognitive Science, like Rumelhart and McClelland, were psychologist by profession.

Empiricist cognitive architecture has gained a voice in every discipline in the cognitive sciences. The increasing acceptance of empiricism is leading not only to the testing of a rapidly-growing number of so-inspired hypotheses but also to a vast reinterpretation of earlier findings in light of radically different postulates. What has been taking place is clearly a Kuhnian paradigm shift. Hence, an exorbitant amount is still to be done. For starters, oddly enough several empiricist researchers are not convinced that their standing agendas are in fact empiricist, that is, that replacing ‘empiricist’ with ‘interactionist’ or with ‘emergentist’ does not black out the ‘empiricist’.

Consider, for example, the book Rethinking Innateness: A Connectionist Perspective on Development  (Elman et al. 1996). After a thorough and outstanding assault of rationalism and defense of empiricism, the group goes on to assert “We are not empiricists” (p. 357). Like many other fearful academics, they view the label ‘empiricist’ as a stigma, not unlike having to bear the Scarlet Letter. It is about time that this stigma be removed, and in that spirit I offer a few clarifications. First, regardless of what Chomsky and Fodor would like us to believe, behaviorism and empiricism are not synonymous, as most versions of connectionism clearly illustrate. Even the simplest neural learning algorithms, such as error backpropagation, offer that which behaviorist could not, statistical means that can carry cognition from learning through finite data to understanding an infinite amount of possibilities. Second, consider the following excerpt—

"We are neither behaviorists nor radical empiricists. We have tried to point out throughout this volume not only that the tabula rasa approach is doomed to failure, but that in reality, all connectionist models have prior constraints of one sort or another. What we reject is representational nativism." (Elman et al. 1996 1996, p. 365)

In Rethinking Innateness, the authors distinguish between three kinds of possible innate constraints: representational, architectural, and chronotopic (timing). A prime example of an architectural constraint is the characteristic 6-layer structure of the human neocortex; for chronotopic constraints, think of embryonic cell migrations. As stated above, the group offers a wealth of innate architectural and chronotopic constraints but reject representational constraints. It is the wealth of mechanisms that can go into delineating what kind of tabula the mind is that leads them to suggest that interactionism entails that empiricism is false. But empiricists have never shunned innateness altogether. The empiricist-rationalist distinction rests squarely on the issue of innate mental representations.

Advancing a strong view of architectural and chronotopic constraints does not depart one from the notion of a tabula rasa. The interaction of the many constraints with the world conforms the tabula—no sane empiricist would ever deny this! —but that does not render the tabula un-rasa, it just delineates what kind of tabula it is (i.e., a nervous system, not a DVD or a 35mm film or an infinite magnetic tape). To put it simply, denying all innate architectural and chronotopic features would be tantamount to claiming the children resemble their parents only because their parents raise them.  No one ever claimed that! The debate between rationalists and empiricists has always been about whether there are certain pieces of knowledge that are represented in the mind that are simply not learned. If you reject representational nativism yet do not reject the existence of something like ideas or mental representations, then you are committed to the tabula rasa, whether you like it or not. It may be unpopular, but it is nevertheless so because rejecting representational nativism without discarding mental representation is affirming that there are no innate ideas. That the type of tabula that it is determines what kind of information can be written on it and that human brains are highly structured does not entail the falsity of empiricism, unless representation is preprogrammed into the slate. Without unlearned representations, a highly structured and complex tabula is as concordant with empiricism as a simple and amorphous pattern-seeking agent.

Clearly, the type of slate that is proposed today is different from what was proposed during the Enlightenment. To Hume, the mind was primarily a passive photocopier of experience; in contrast, current neural networks are much more active in their assimilation of environmental information. Moreover, while Hume thought that that human minds associate the compiled copies of experience according to three domain-general types of association, connectionist neural networks are universal approximators that modularize as functional approximations consolidate because of the details of the surrounding environment and, therefore, in consequence, these readily develop mechanisms that go beyond association through association itself (see How You Know What You Know for a review). Advancing a stronger, more complex view of the cognitive slate does not distance the account from empiricism since it rejects representational nativism, just like Elman et al. 1996 did.

It is telling that connectionists naturally gravitate toward empiricism in spite of the stigma surrounding the tradition and even their own explicit assertions and roundabout philosophical identifications. Ultimately, the hallmark dispute among connectionist and classicalists is the question of what kind of tabula the mind is, a question that does not directly concern the rationalist/empiricist distinction but results from it by entailment. It is really just a practical matter that, whereas syntactic or logical engines require innate representations, complex neuronal slates like ours do not. Then again, it is also a practical matter that the only intelligent beings we know of are born with highly complex neural networks. Deep down, I am inclined to think that Fodor’s Informational Atomism is logically correct—if the mind works like a logical or syntactic engine, then all simple concepts must be innate. As Barsalou (1999) notes, there are no accounts on offer for how simple symbols can be acquired by a classical cognitive architecture or any logical or syntactic engine, and this may very well be because there are no possible accounts at all. This admission, however, should not lead us to accept Fodor’s theory of concept, but rather it should convince us that the mind is not a Turing machine (like the image below) or a syntactic engine (cf., Pinker 2005).

As the evidence mounts, even Chomsky had to abandon most of the original postulates of generative linguistics, including the important distinction between surface structure and deep structure and also the view that syntax is a totally autonomous faculty that does not derive or associate at all with the lexicon.  The Minimalist Program (1995) reduced the philosophical rationalism of Chomsky's theory to such an extent that several academics that have based their own work on generative models, suddenly finding themselves in a theoretical void that threatens to undermine their research, have chosen either to ignore it entirely or to attempt to undermine the program.  But this is just one example of how rationalist philosophy of mind is undergoing its slow death, weakening as data piles up.  As the first generation of cognitive scientists dies out and the third generation starts to assume positions of power, the stigma branded upon empiricism will weaken.  The likely result is a renewal that will allow funding to flow to new experimental techniques and to innovative practical application across the interrelated disciplines.  Exciting times lie ahead.



- Anderson, J.R. (1982). “Acquisition of cognitive skill”. Psychological Review 89: 369-406.
- Arnauld, A. & Lancelot, C. (1660). General and Rational Grammar: The Port-Royal Grammar. J. Rieux and B.E. Rollin (trans.). The Hague: Mouton, 1975.
- Arnauld, A. & Nicole, P. (1662). Logic, or The Art of Thinking; being The Port-Royal Logic. Thomas Spencer Baynes (trans.). Edinburgh: Sutherland and Knox, 1850.
- Barsalou, L.W. (1999). “Perceptual symbol systems.” Behavioral and Brain Sciences, 22: 577-609.
- Bechtel, W., Abrahamsen, A. & Graham, G. (1998). "The Life of Cognitive Science". A Companion to Cognitive Science. W. Bechtel & G. Graham (eds.). Massachusetts: Blackwell Publishers Ltd.
- Boole, G. (1854). An Investigation of the Laws of Thought on Which are Founded the Mathematical Theories of Logic and Probabilities. London: Macmillan.
- Brooks, R.A. (1991). “Intelligence Without Representation.” Artificial Intelligence Journal 47: 139–160.
- Chomsky, N. (1957). Syntactic Structures. New York: Mouton de Gruyter.
- Chomsky, N. (1959). "A Review of B. F. Skinner's Verbal Behavior." Language, 35, No. 1: 26-58.
- Chomsky, N. (1966). Cartesian Linguistics: A Chapter in the History of Rationalist Thought. New York: Harper & Row.
- Chomsky, N. (1967). “Preface to the 1967 reprint of ‘A Review of Skinner's Verbal Behavior’.” Readings in the Psychology of Language. Leon A. Jakobovits & Murray S. Miron (eds.). Prentice-Hall, Inc. pp. 142-143.
- Chomsky, N. (1995). The Minimalist Program. Cambridge, MA: MIT Press.
- Dreyfus, H.L. (1992). What Computers Still Can’t Do: A Critique of Artificial Reason. Cambridge, MA: MIT Press.
- Elman, J. L. (1998). “Connectionism, artificial life, and dynamical systems: New approaches to old questions.” A Companion to Cognitive Science. W. Bechtel & G. Graham (eds.) Oxford: Basil Blackwood.
- Elman, J.L., Bates, E.A., Johnson, M.H., Karmiloff-Smith, A., Parisi, D., Plunkett, K. (1996). Rethinking Innateness: A Connectionist Perspective on Development. Cambridge, MASS: MIT Press.
- Feldman, J.A. (1981). “A connectionist model of visual memory.” Parallel models of associative memory. G.E. Hinton y J.A. Anderson (eds.). Nueva Jersey: Erlbaum.
- Fodor, J.A. (2003). Hume Variations. New York: Oxford University Press.
- Fodor, J.A. & Pylyshyn, Z.W. (1988). “Connectionism and Cognitive Architecture: A Critical Analysis.” Cognition 28: 3-71.
- Goldberg, A.E. (2004). “But do we need Universal Grammar? Comment on Lidz et al.”(2003)” Cognition 94: 77-84.
- Hare, M. & Elman, J.L. (1995). “Learning and morphological change.” Cognition 56: 61-98.
- Haugeland, J. (ed.) (1985). Artificial Intelligence: The Very Idea. Cambridge, MA: MIT Press.
- Hazlehurst, B. & Hutchins, E. (1998). “The emergence of propositions from the co-ordination of talk and action in a shared world.” Language and Cognitive Processes 13(2/3): 373-424.
- Hebb, D. (1949). The Organization of Behavior: A Neuropsychological theory. New York: Wiley.
- Hutchins, E. & Hazlehurst, B. (1995). “How to invent a lexicon: the development of shared symbols in interaction.” Artificial Societies: the computer simulation of social life. N. Gilbert & R. Conte (eds.). London: UCL Press. pp. 157-189.
- Kuhn, T. (1962). The Structure of Scientific Revolutions. Chicago: University of Chicago Press, 1970. (2nd revised edition)
- Lakoff, G. (1987). Women, Fire, and Dangerous Things: What Categories Reveal About the Mind. Chicago: The University of Chicago Press.
- Langacker, R.W. (1987). Foundations of Cognitive Grammar. Stanford, CA: Stanford University Press.
- Lewis, J.D., & Elman, J.L. (2001). “Learnability and the statistical structure of language: Poverty of stimulus arguments revisited.” Proceedings of the 26th Annual Boston University Conference on Language Development.
- MacWhinney, B. (2013). “The Logic of a Unified Model”. S. Gass and A. Mackey (eds.). Handbook of Second Language Acquisition. New York: Routledge. pp. 211-227.
- McClelland, J.L. & Rumelhart, D.E. (1981). “An interactive activation model of context effects in letter perception: Part 1. An account of basic findings.” Psychological Review 88: 375-407.
- McClelland, J.L. (1989). “Parallel distributed processing: Implications for cognition and development.” Morris, R. (ed.) Parallel distributed processing: Implications for psychology and neurobiology. New York: Oxford University Press.
- McCulloch, W.S. & Pitts, W. (1943). “A logical calculus of the ideas immanent in nervous activity.” Bulletin of Mathematical Biophysics 5: 115–137.
- Newell, A., Shaw, J.C. & Simon, H.A. (1959). “Report on a general problem-solving program”. Proceedings of the International Conference on Information Processing . pp. 256-264.
- Pinker, S. (2005). "So How Does The Mind Work?" Mind and Language 20, 1: 1-24.
- Prince, A. & Smolensky, P. (1997). “Optimality: From Neural Networks to Universal Grammar”. Science 275: 1604-1610.
- Prinz, J.J. (2002). Furnishing the Mind. Massachusetts: MIT Press.
- Quine, W.V.O. (1960). Word and Object. Massachusetts: MIT Press.
- Rochester, N., Holland, J.H., Haibt, L.H., & Duda, W.L. (1956). “Tests on a cell assembly theory of the action of the brain, using a large digital computer.” IRE Transactions on Information Theory 2: 80-93.
- Rosenblatt, F. (1962). Principals of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. Washington, D.C.: Spartan Books.
- Skinner, B.F. (1957). Verbal Behavior. Acton, MA: Copley, 1992.
- Thelen, E. & Bates, E. (2003). “Connectionism and dynamic systems: are they really different?” Developmental Science 6, 4: 378-391.
- Tomasello, M. (2006). “Acquiring linguistic constructions”. Handbook of Child Psychology. Kuhn, D. & Siegler, R. (eds.). New York: Wiley.
- Turing, A.M. (1936). "On Computable Numbers, with an Application to the Entscheidungsproblem". Proceedings of the London Mathematical Society, 2, 42: pp. 230–65, 1937.
- van Gelder, T.J. (1999). “Defending the dynamical hypothesis.” Dynamics, Synergetics, Autonomous Agents: Nonlinear Systems Approaches to Cognitive Psychology and Cognitive Science. W. Tschacher & J.P. Dauwalder (eds.) Singapore: World Scientific. pp. 13-28.
- von Neumann, J. (1945). "First Draft of a Report on the EDVAC". Originally confidential [property of the United States Army Ordnance Department].
- Zwaan, R.A. (2004). “The Immersed Experiencer: Toward an embodied theory of language comprehension.” The Psychology of Learning and Motivation 44: 35-62.

If you enjoyed this article, you may also like: