Connectionism: Debates in Psychological Explanation - Preface
MacDonald (Cynthia) & MacDonald (Graham)
Source: MacDonald (Cynthia) & MacDonald (Graham) - Connectionism: Debates in Psychological Explanation - Vol. 2
Paper - Abstract

Paper StatisticsNotes Citing this PaperColour-ConventionsDisclaimer

Preface (Full Text)

  1. Those involved in the study of human behaviour and cognition have been interested for decades in the question of what an integrated overall psychology might look like. We are all familiar with the kinds of psychological states that common-sense or ‘folk' psychology recognizes. On a day-to-day basis, we make sense of, or explain, human behaviour by attributing to people beliefs, desires and the like — states with propositional content. My going to the refrigerator and reaching for a coke is explained by the fact, among others, that I desire a coke and I believe that there is a coke in the fridge; states which jointly cause my behaviour and do so in virtue of their propositional contents. If my belief and desire had had different contents — if, for example, they had concerned, not coke, but cigarettes — I would have behaved differently: rather than go to the refrigerator, I would have headed for my coat pocket.
  2. We know that many ‘folk' theories, such as ‘folk' physics, are often left behind by the sciences that they found. We also know, however, that beliefs, desires and other propositional states currently figure in explanations in various parts of psychology. For instance, propositional states figure in explanations of various aspects of cognition, for example, memory, and in explanations in behavioural psychology: human behaviour is thereby usefully predicted and explained. This raises the question of what kinds of descriptions a completed psychology will employ in its explanations. Will that science find room for descriptions of beliefs and desires, states with propositional content? And if so, how might these descriptions relate to descriptions of cognitive processes, the psychological mechanisms involved, for example, in information storage and retrieval, and descriptions of the neurophysiological structures and connections in the brain that underlie cognition?
  3. Since the work of David Marr and others, it has become customary for those working in the behavioural sciences to expect that an adequate overall psychology will employ descriptions at many levels of explanation, and that these explanatory levels will supplement rather than compete with one another. In his seminal work, Vision, Marr (19821) developed a theory of visual perception that employs three distinct levels of description. The first, computational, level describes the function carried out by the visual system, whereas the second, algorithmic, level describes how the system carries out that function, how it computes it. The third, implementational, level describes the physical means by which the function is carried out, how the function is physically realized.
  4. Consider, for example, an ADDER, a system that takes numbers as arguments and delivers sums as values. We know that there are many types or kinds of physical mechanisms that qualify as ADDERS, even though they are constructed physically very differently. Take the function z = 3 (x + y). An electronic calculator will realize this function physically differently from a human being; but both are capable of carrying out the function. So there is a difference between a description of an ADDER at the computational level and a description of it at the level of physical implementation. But there is another way in which ADDERS can differ, a way that emerges at the algorithmic level of description. Here differences in the ways that various ADDERS carry out the same function at the computational level, differences in the algorithms employed by the systems which need not imply differences at the level of physical implementation, may emerge. Two ADDERS may compute the function z = 3 (x + y) differently. The first may compute the function by adding the values of x and y and multiplying the result by 3, whereas the second may multiply each of the values of x and y by 3 and add the results. These different step-by-step procedures are different algorithms.
  5. Marr's three-level model is widely accepted by theorists involved in the study of human behaviour and cognition. The model is a particularly useful one for illustrating how psychology might both accommodate different varieties of description and explanation and show how these varieties relate to and integrate with one another. It naturally suggests a model for human cognition which can integrate descriptions of propositional states at the computational level with descriptions of states at other (cognitive processing and neurophysiological) explanatory levels.
  6. Theories of cognition concern themselves mainly with the first two of Marr's three levels, the computational and the algorithmic levels. Until recently, such theories have been dominated by a particular view of cognition known as the Classical view (see Fodor and McLaughlin in chapter 52 of this volume3). This view, which has its source in the work of Alan Turing, construes cognition as symbol manipulation. The idea is that cognitive processes involve manipulating objects that have a syntax and a semantics4, and which can be combined and transformed according to fixed rules. Such symbols are also capable of being stored in and retrieved from memory. In short, the Classical view construes thought as language-like in involving the manipulation of symbols, or representations, that have both a syntax and a semantics5.
  7. However, in the 1980s another, alternative view of cognition, based on a model known as connectionist, or parallel distributed processing, emerged. A motivating factor in the development of this new conception was to construct a model of cognition that more closely resembles the structure of the human brain, and so is more biologically plausible than classical models. What prompted the conception was the recognition that the brain is a neural network. Whereas classical models of cognition work serially, i.e. by creating, comparing and transforming strings of symbols according to fixed rules which are seen as governing cognition itself, connectionism offers a very different conception of the basic processing involved in cognition. Connectionist models do not work serially but in parallel and interactive ways. The models involve a network of elementary units or nodes, each of which has a degree of activation. These units are connected to each other and the connections are weighted, so that, given the degree of activation of a unit and the weight of its connections to others, it will either excite or inhibit those other units. The behaviour of the system as a whole is determined by the initial degree of activation of the units plus the strengths of the weights connecting them. Crucially, in widely distributed networks, the individual units or nodes are not semantically interpretable, and the semantically interpretable units, if there are any, do not correspond in a one-to-one way with any discrete unit or collection thereof. So these models represent a departure from the Classical view of cognition as symbol manipulation.
  8. The emergence of connectionist models of cognition has created a great deal of controversy among those involved in the study of human behaviour and cognition. Two important debates have emerged from the controversy, which form the subject matter of this second volume of debates on psychological explanation. (The first volume6, Philosophy of Psychology: Debates on Psychological Explanation, is also published by Basil Blackwell.) The first centres on the assumption that three crucial and essential features of human cognition, or thought, are
    • (a) its productivity,
    • (b) its inferential coherence, and
    • (c) its systematicity;
    and that any adequate theory of cognition must explain these features. The claim made by opponents of connectionist models is that they do not explain these features; in particular, they do not explain the systematicity of thought. Thoughts, or cognitive capacities, are systematic if and only if they are intrinsically connected, according to Jerry Fodor and Zenon Pylyshyn (see chapter 37). Thought is systematic in the sense that the ability to think aRb is intrinsically connected to the ability to think bRa. (Classical models are capable of explaining systematicity because they construe cognitive processes as operating on symbols; constituents that can be moved around and transformed in strings according to rules. But, it is argued, because connectionist models do not conceive of cognition as involving the manipulation of symbols, they do not explain systematicity in this way. This presents connectionists with a challenge: if they cannot explain systematicity in a Classical way, can they explain it at all?
  9. The response is that the proper understanding of connectionism is one according to which, although the individual activation values of units are incapable of semantic interpretation, patterns of activity, or vectors of activity, are capable of such interpretation. The result, it is claimed, is that connectionist models are able to explain the systematicity of thought, albeit in a non-Classical way.
  10. Both sides of this controversy assume that descriptions of propositional states such as beliefs and desires, the states of common-sense psychology, will figure in an overall theory of cognition. The issue is whether connectionism is up to the job of explaining various features of these states. One side argues that propositional states have certain features, connectionist models do not have these features, therefore connectionist models are not adequate to the task of providing an adequate overall theory of cognition.
  11. This way of viewing the first debate in the volume, that between classicism and connectionism, links it to the second, that between connectionism and eliminativism. The core of the first debate centres on the issue of whether connectionism is compatible with common-sense, or propositional attitude, psychology. If not, and if one believes that descriptions of propositional attitudes will figure in an overall theory of cognition, then connectionism is under threat. The second debate too assumes that propositional states have certain features, viz.
    • (a) semantic evaluability,
    • (b) functional discreteness, and
    • (c) causal efficacy.
    The question that arises, and upon which the debate focuses, is whether the states of certain connectionist models of memory (and other propositional states) have these three features. One side argues that they do not, so that if these models are correct, then common-sense psychology is false. It further argues that if common-sense psychology is false, then there are no propositional states (i.e. that eliminativism is true). The response to this argument is that the first conditional in it is false: connectionist models have all, or most, of the features ascribed to the states of common-sense psychology.
  12. Putting these debates side by side raises interesting and important philosophical questions about the nature of propositional states and methodological questions concerning the possible conflict between emerging scientific theories of the mind and common-sense conceptions. The philosophical questions concern the appropriate way to characterize propositional attitudes. It is indicative of the unsettled nature of discussions of the common-sense conception that there is no overlap in the two debates presented in this volume in the features taken to be characteristic of and essential to propositional states. This in itself raises an interesting methodological question about the appropriate way to respond to a potential conflict between common-sense conceptions and emerging scientific paradigms of the mind. The authors in the second debate accept that if there is a conflict between the emerging connectionist paradigm and the common-sense conception of propositional states, it is the common-sense conception that is under threat. The authors in the first debate, on the other hand, accept that if there is a conflict between the connectionist paradigm and the common-sense conception, it is the connectionist paradigm that is under threat. Neither of the two debates questions the conception of propositional states appealed to in the controversy, and so neither exploits the strategy of attempting to reconcile emerging connectionist paradigms with the common-sense conception by questioning the paradigm of the common sense conception appealed to. However, by appealing to different conceptions of propositional attitudes, and by invoking different methodological strategies, the two debates in this volume should stimulate discussion on these important issues.
  13. Connectionist architectures promise to provide novel and important insights into human behaviour and cognition, but the development of the theory of such architectures is still in its infancy. One point that emerges from both debates in the present volume is the unsettled nature of opinion about exactly how connectionist models are to be understood, and in what way they differ from classical models of cognition. The debates should also stimulate discussion on this important issue.
  14. Each debate consists of papers previously published which present seminal views on the relation between connectionist architectures and propositional states, along with entirely new contributions by certain authors which further the debate. Introductions to each debate are also included. These introductions aim to provide assistance to the reader in identifying the central issues in dispute, and make the volume valuable for upper level courses in philosophy of psychology and cognitive science, whether these be in upper-level undergraduate classes or at postgraduate level. Students of cognitive science, psychology and philosophy will stand to benefit from such a course.
  15. We would like to thank … [… snip …]

In-Page Footnotes

Footnote 1: See "Marr (David) - A Theory of Vision".

Footnote 2: See "Fodor (Jerry) - Connectionism and the Problem of Systematicity (Continued): Why Smolensky's Solution Still Doesn't Work".

Footnote 3: In this volume, we adopt the following stylistic convention regarding Cassical/Connectionist theory and classicism/connectionism. All uses of the terms related to the theory employ an upper-case ‘C'; all other uses employ a lower-case ‘c'.

Footnote 6: See "MacDonald (Cynthia) & MacDonald (Graham), Eds. - The Philosophy of Psychology: Debates in Psychological Explanation - Vol. 1".

Footnote 7: See "Fodor (Jerry) & Pylyshyn (Zenon) - Connectionism and Cognitive Architecture: A Critical Analysis".

Text Colour Conventions (see disclaimer)

  1. Blue: Text by me; © Theo Todman, 2020
  2. Mauve: Text by correspondent(s) or other author(s); © the author(s)

© Theo Todman, June 2007 - Nov 2020. Please address any comments on this page to File output:
Website Maintenance Dashboard
Return to Top of this Page Return to Theo Todman's Philosophy Page Return to Theo Todman's Home Page