Skip to main content
    • Aa
    • Aa

Responsibility and Decision Making in the Era of Neural Networks*

  • William Bechtel (a1)

Many of the mathematicians and scientists who guided the development of digital computers in the late 1940s, such as Alan Turing and John von Neumann, saw these new devices not just as tools for calculation but as devices that might employ the same principles as are exhibited in rational human thought. Thus, a subfield of what came to be called computer science assumed the label artificial intelligence (AI). The idea of building artificial systems which could exhibit intelligent behavior comparable to that of humans (which could, e.g., recognize objects, solve problems, formulate and implement plans, etc.) was a heady prospect, and the claims made on behalf of AI during the 1950s and 1960s were impossibly ambitious (e.g., having a computer capture the world chess championship within a decade). Despite some theoretical and applied successes within the field, serious problems soon became evident (of which the most notorious is the frame problem, which involves the difficulty in determining which information about the environment must be changed and which must be kept constant in the face of new information). Instead of fulfilling the goal of quickly producing artificial intelligent agents which could compete with or outperform human beings, by the 1970s and 1980s AI had settled into a pattern of slower but real progress in modeling or simulating aspects of human intelligence. (Examples of the advances made during this period were the development of higher-level structures for encoding information, such as frames or scripts, which were superior to simple prepositional encodings in supporting reasoning or the understanding of natural [as opposed to computer or other artificial] language texts, and the development of procedures for storing information about previously encountered cases and invoking these cases in solving new problems.)

Hide All

1 The term “artificial intelligence” was apparently invented by John McCarthy in the context of a seminal conference at Dartmouth College in 1956. For accounts of the early history of AI, see McCorduck Pamela, Machines Who Think (San Francisco: W. H. Freeman, 1979); and Gardner Howard, The Mind's New Science: A History of the Cognitive Revolution (New York: Basic Books, 1985).

2 McCarthy John and Hayes Patrick J., “Some Philosophical Problems from the Standpoint of Artificial Intelligence,” in Machine Intelligence, ed. Meltzer B. and Michie D. (Edinburgh: Edinburgh University Press, 1969), pp. 463502. For a fairly recent review of work on the frame problem, see Ford K. M. and Hayes P. J., eds., Reasoning Agents in a Dynamic World: The Frame Problem (London: JAI Press, 1991).

3 Schank Roger C. and Abelson Robert P., Scripts, Plans, Goals, and Understanding (Hillsdale, NJ: Lawrence Erlbaum, 1977); and Minsky Marvin, “A Framework for Representing Knowledge,” in The Psychology of Human Vision, ed. Winston P. H. (New York: McGraw-Hill, 1975).

4 Schank Roger C. and Reisbeck Christopher K., Inside Case-based Reasoning (Hillsdale, NJ: Lawrence Erlbaum, 1989).

5 Buchanan Bruce G. and Shortliffe Edward H., eds., Rule-based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project (Reading, MA: Addison-Wesley, 1984).

6 One may question whether turning off an AI system really constitutes punishment. This depends in part upon whether it makes sense to construe AI systems as having interests. Designing systems that have something recognizable as interests is now an active area of AI research, especially in the field of artificial life. Exploring this topic, however, is beyond the scope of this essay. For now I will simply assume that AI research will continue to be successful in developing artificial agents that resemble humans in their decision making, including possession of motivational states.

7 Larry May, in personal discussion, has proposed an alternative perspective. When a human agent becomes intoxicated and begins to behave in irresponsible ways that are difficult to predict, we do not absolve the agent, but hold him or her responsible for becoming intoxicated. Perhaps we should hold AI designers who devise agents which they know will behave in ways they cannot foresee, similarly responsible. The context of constructing AI systems is, I think, significantly different from the context of becoming intoxicated. The systems are created because it is anticipated that they will generate many good outcomes (solving problems better than humans, etc.). The better analogy, I contend, is with giving birth to children. One hopes that one's children will be agents of good, even if, in ways that are currently unpredictable, some will cause great harm. We do not hold parents morally responsible for the actions of their (adult) offspring if they have done their best to provide an appropriate upbringing; neither should we hold AI designers responsible for the artificial agents they create if they have taken due precautions.

8 See Holland John H., Holyoak Keith J., Nisbett Richard E., and Thagard Paul R., Induction: Processes of Inference, Learning, and Discovery (Cambridge, MA: MIT Press, 1986).

9 During the 1950s and 1960s, a number of researchers actively pursued the alternative connectionist or neural network strategy for creating artificial intelligent agents. See Rosenblatt Frank, The Principles of Neurodynamics (New York: Spartan, 1962); and Selfridge Oliver G. and Neisser Ulric, “Pattern Recognition by Machine,” Scientific American, vol. 203 (08 1960), pp. 6068. A critical analysis of the limitations of such systems by Minsky Marvin and Papert Seymour, Perceptrons (Cambridge, MA: MIT Press, 1969), helped to reduce interest in this approach. For a humorous recounting of the fall and reemergence of the neural network alternative, see Papert Seymour, “One AI or Many?Daedalus, vol. 117 (1988), pp. 114.

10 Perhaps the seminal event in the reemergence of such models was the publication of Rumelhart David E., McClelland James L., and the PDF Research Group, Parallel Distributed Processing: Explorations in the Microstructure of Cognition, vol. 1: Foundations (Cambridge, MA: MIT Press, 1986); and McClelland James L., Rumelhart David E., and the PDP Research Group, Parallel Distributed Processing: Explorations in the Microstructure of Cognition, vol. 2: Psychological and Biological Models (Cambridge, MA: MIT Press, 1986). As their titles indicate, Rumelhart and McClelland's preferred term for these models is parallel distributed processing models. For an elementary exposition of these models and their application to modeling cognition, see Bechtel William and Abrahamsen Adele A., Connectionism and the Mind: An Introduction to Parallel Processing in Networks (Oxford: Basil Blackwell, 1991).

11 Bechtel William, “Attributing Responsibility to Computer Systems,” Metaphilosophy, vol. 16 (1984), pp. 296306.

12 Dennett Daniel C., “Intentional Systems,” Journal of Philosophy, vol. 68 (1971), pp. 87106.

13 Franz Brentano introduced the use of the term “intentionality” to refer to the fact that mental states are about things. He emphasized that this relation was unlike ordinary extensional relations since one could have mental states whose content did not exist. See Brentano Franz, Psychology from an Empirical Standpoint [1874], trans. Rancurello A. C., Terrell D. B., and McAlister L. L. (New York: Humanities Press, 1973).

14 See Fodor Jerry A., The Language of Thought (New York: Crowell, 1975), and Psychosemantics: The Problem of Meaning in the Philosophy of Mind (Cambridge, MA: MIT Press, 1987).

15 Searle John R., “Minds, Brains, and Programs,” Behavioral and Brain Sciences, vol. 3 (1980) pp. 417–24. I have argued elsewhere that Searle's argument that we, and not AI systems (even those with robotic bodies that seem to satisfy the conditions I set out below), enjoy intrinsic intentionality rests on an illusion stemming from our use of language: with language we acquire the possibility of meta-representations that allow us to characterize the content of our mental states. This makes it seem as if, whenever we are in a mental state, its contents are directly presented to us and we know directly what they are about (part of Searle's argument that we and not AI systems have intrinsic intentionality). I contend, rather, that we have more than one representational system, and are able to use one to specify the contents of the other. See Bechtel William and Abrahamsen Adele A., “Connectionism and the Future of Folk Psychology,” in Natural and Artificial Minds, ed. Burton Robert G. (Albany, NY: SUNY Press, 1993), pp. 69100.

16 The origins of eliminative materialism are found in the work of philosophers such as Feyerabend Paul K., “Materialism and the Mind-Body Problem,” Review of Metaphysics, vol. 17 (1963), pp. 4967; and Rorty Richard, “Mind-Body Identity, Privacy, and Categories,” Review of Metaphysics, vol. 19 (1965), pp. 2454. The most vociferous contemporary statements of eliminative materialism are by Churchland Patricia S., Neurophilosophy: Towards a Unified Science of the Mind-Brain (Cambridge, MA: MIT Press, 1986); and Churchland Paul M., A Neurocomputational Perspective: The Nature of Mind and the Structure of Science (Cambridge, MA: MIT Press, 1989). A variant of the eliminative materialist position is found in Stich Stephen P., From Folk Psychology to Cognitive Science (Cambridge, MA: MIT Press, 1983).

17 For a review of the philosophical arguments for and against functionalism, see Bechtel William, Philosophy of Mind: An Overview for Cognitive Science (Hillsdale, NJ: Lawrence Erlbaum, 1988).

18 For contemporary reviews of cognitive psychology, see Barsalou Lawrence, Cognitive Psychology (Hillsdale, NJ: Lawrence Erlbaum, 1994); and Anderson John R., Cognitive Psychology and Its Applications, 3d ed. (San Francisco: Freeman, 1990).

19 Pylyshyn Zenon W., Computation and Cognition: Toward a Foundation for Cognitive Science (Cambridge, MA: MIT Press, 1984).

20 Newell Allen and Simon Herbert A., Human Problem Solving (Englewood Cliffs, NJ: Prentice-Hall, 1972).

21 Langley Patrick, Simon Herbert A., Bradshaw Gary L., and Zytkow J. M., Scientific Discovery: Computational Explorations of the Creative Processes (Cambridge, MA: MIT Press, 1987).

22 For a detailed development of this claim that does not itself invoke neural networks, see Margolis Howard, Patterns, Thinking, and Cognition (Chicago: University of Chicago Press, 1987).

23 This network is described more fully in Bechtel William, “Natural Deduction in Connectionist Systems,” Synthese, vol. 101, no. 3 (1994), pp. 433–63.

24 This distinction is due to Ryle Gilbert, The Concept of Mind (New York: Barnes and Noble, 1949).

25 Ramsey William, Stich Stephen P., and Garon J., “Connectionism, Eliminativism, and the Future of Folk Psychology,” Philosophical Perspectives, vol. 4 (1990), pp. 499533.

26 Hinton Geoffrey E., “Learning Distributed Representations of Concepts,” Proceedings of the Eighth Annual Conference of the Cognitive Science Society (Hillsdale, NJ: Lawrence Erlbaum, 1986), pp. 112; and van Gelder Timothy, “What is the ‘D’ in ‘PDP’? A Survey of the Concept of Distribution,” in Philosophy and Connectionist Theory, ed. Ramsey William, Stich Stephen P., and Rumelhart David E. (Hillsdale, NJ: Lawrence Erlbaum, 1991), pp. 3359.

27 Churchland, A Neurocomputational Perspective (supra note 16).

28 For the initial challenge to the classical view of concepts, see Rosch Eleanor and Mervis Carolyn B., “Family Resemblances: Studies in the Internal Structure of Categories,” Cognitive Psychology, vol. 7 (1975), pp. 573605. Rosch and Mervis focused primarily on establishing that categories had a “prototype structure” (that is, that members of a category were ranked from more prototypical to less prototypical), not on the mechanism by which people made assignments to categories. For a review of subsequent psychological research which has construed comparison to prototypes as the basis of categorization (as well as an alternative view which construes comparison to multiple exemplars or actual members of a category as the basis), see Barsalou, Cognitive Psychology (supra note 18); and Smith Edward E., “Categorization,” in Invitation to Cognitive Science, vol. 3: Thinking, ed. Osherson Daniel N. and Lasnik Howard (Cambridge, MA: MIT Press, 1990), pp. 3353.

29 Churchland Paul M., “The Neural Representation of the Social World,” in Minds and Morals, ed. May L., Friedman M., and Clark A. (Cambridge, MA: MIT Press, 1995).

30 See Bechtel and Abrahamsen, “Connectionism and the Future of Folk Psychology” (supra note 15).

31 For an explication of logical behaviorism, see Bechtel, Philosophy of Mind (supra note 17), ch. 2.

32 For an example, see Bechtel William and Richardson Robert C., Discovering Complexity: Decomposition and Localization as Strategies in Scientific Research (Princeton, NJ: Princeton University Press, 1993), ch. 8.

33 See Rumelhart David E. and McClelland James L., “On Learning the Past Tense of English Verbs,” in Parallel Distributed Processing, vol. 2 (supra note 10), ch. 18; and Plunkett Kim and Marchman Virginia, “U-shaped Learning and Frequency Effects in a Multilayered Perceptron,” Cognition, vol. 38 (1991), pp. 160.

34 See Hinton Geoffrey E. and Shallice Timothy, “Lesioning an Attractor Network: Investigations of Acquired Dyslexia,” Psychological Review, vol. 98 (1991), pp. 7495.

35 See John Mark F. St. and McClelland James L., “Learning and Applying Contextual Constraints in Sentence Comprehension,” Artificial Intelligence, vol. 46 (1990), pp. 217–57; and Miikkulainen Risto, Subsymbolic Natural Language Processing: An Integrated Model of Scripts, Lexicon, and Memory (Cambridge, MA: MIT Press, 1993).

36 This approach of embodying neural networks in robots has been pursued by several researchers. In some of the most interesting work, it has been coupled with a procedure for evolving new network architectures through application of the genetic algorithm. (The genetic algorithm is a procedure for revising computer code by a process of random variation and selective retention of improved variants.) See Nolfi S., Elman Jeffrey L., and Parisi D., “Learning and Evolution in Neural Networks,” Adaptive Behavior, vol. 3, no. 1 (1994), pp. 528; and Nolfi S., Miglino O., and Parisi D., “Phenotypic Plasticity in Evolving Neural Networks,” Proceedings of the First Conference from Perception to Action, ed. Gaussier D. P. and Nicoud J. D. (Los Alamitos, CA: IEEE Press, 1994). I have argued for the importance of such approaches in creating networks with genuine intentionality in Bechtel William, “The Case for Connectionism,” Philosophical Studies, vol. 71 (1993), pp. 119–54.

37 See Fetzer James, Artificial Intelligence: Its Scope and Limits (Dordrecht: Kluwer, 1990); and Peirce Charles Sanders, “Speculative Grammar,” in Hartshorne Charles and Weiss Paul, eds., Collected Papers of Charles Sanders Peirce, vol. 2, Elements of Logic (Cambridge: Harvard University Press, 1960).

38 Savage-Rumbaugh E. Sue, Ape Language: From Conditioned Response to Symbol (New York: Columbia University Press, 1986).

39 The importance of these contrast relations is stressed by Deacon Terrence W., Symbolic Origins (New York: W. W. Norton, forthcoming).

40 This process is known as protocol analysis. See Ericsson K. Anders and Simon Herbert A., Protocol Analysis: Verbal Reports as Data (Cambridge, MA: MIT Press, 1984).

41 See Bechtel and Richardson, Discovering Complexity (supra note 32).

42 For a discussion of the relative merits of cluster analysis and principal-components analysis, and a detailed example of using principal-components analysis to understand the behavior of a network, see Elman Jeffrey L., “Distributed Representations, Simple Recurrent Networks, and Grammatical Structure,” Machine Learning, vol. 7 (1991), pp. 195225.

43 Davidson Donald, “Rational Animals,” Dialectica, vol. 36 (1982), pp. 318–27.

44 See Bechtel William, “Biological and Social Constraints on Cognitive Processing: The Need for Dynamical Interactions between Levels of Inquiry,” Canadian Journal of Philosophy, Supplementary Volume 20 (1994), pp. 133–64.

45 Nisbett Richard E. and Wilson Timothy D., “Telling More Than We Can Know: Verbal Reports on Mental Processes,” Psychological Review, vol. 84 (1974), pp. 231–59.

46 Vygotsky Lev S., Thought and Language (Cambridge, MA: MIT Press, 1934, 1962).

47 Newell and Simon, Human Problem Solving (supra note 20). See also Ericsson and Simon, Protocol Analysis (supra note 40).

* This essay was partly prepared while I was a visiting scholar of the Experimenteel-psychologische Onderzoekschool and affiliated with the Theoretical Psychology Faculty in the Department of Psychonomics at the Vrije Universiteit, Amsterdam. I am very grateful for the support I received there, especially conversations relevant to this essay with Huib Looren de Jong. I also thank Andy Clark, Larry May, and the editor of this volume, Ellen Paul, for their comments and suggestions on previous drafts of this essay.

Recommend this journal

Email your librarian or administrator to recommend adding this journal to your organisation's collection.

Social Philosophy and Policy
  • ISSN: 0265-0525
  • EISSN: 1471-6437
  • URL: /core/journals/social-philosophy-and-policy
Please enter your name
Please enter a valid email address
Who would you like to send this to? *


Full text views

Total number of HTML views: 0
Total number of PDF views: 1 *
Loading metrics...

Abstract views

Total abstract views: 102 *
Loading metrics...

* Views captured on Cambridge Core between September 2016 - 24th October 2017. This data will be updated every 24 hours.