A Critical Review of What Computers Still Can't Do, by Hubert Dreyfus (from 1995)

 Ron Barnette, Valdosta State University


Review of What Computers Still Can't Do: a Critique of Artificial Reason, by Hubert L. Dreyfus. 1993. The MIT Press: Cambridge, Massachusetts and London, England. 

Reviewer: Ron Barnette, Professor of Philosophy (Ph.D. University of California, Irvine), Valdosta State University, University System of Georgia, USA.


This book bears a new title (adding the word 'Still') for two earlier editions (1972, 1979) of a philosophical work aimed at undermining a significant physicalistic theory of mind: Classical Cognitivism, central to nearly all (until recently) approaches to Artificial Intelligence (AI) models of human intelligence, inspired by the directions set forth by Turing and von Neumann. What is of particular interest about the new 1993 issue is the thirty-eight page Introduction, which articulates Dreyfus' scope of resistance to the mind/machine research paradigm; what follows the new Introduction are seven pages of Notes, and Introductions to the second and first editions, together with the original text of What Computers Can't Do( Introduction material). As such, I will focus my remarks on the latest Introduction, which reiterates and attempts to strengthen earlier arguments in the up-to-date context of new directions in AI research and design.
In a tone of almost "I told you so---but why won't it go away?!," Dreyfus tries to lay to final rest the AI approach dubbed by John Haugeland (1985) 'Good Old Fashioned Artificial Intelligence,' or GOFAI. Inspired by a Turing model of intelligent behavior as essentially computational, the thesis of GOFAI is that the processes underlying intelligence are symbolic in nature. More specifically: GOFAI models human intelligence as von Neumann computational architectures that (1) perform computations on abstract symbolic representations, (2) by computations governed by a stored program that contains an explicit list of instructions or rules which transform these symbolic representations into new symbolic states, and (3) in terms of which these computations are performed in serial fashion by a CPU that has information stored in the computers' permanent memory. As such, GOFAI depicts mentality within the context of what philosophers know as the Representational Theory of Mind, according to which the mind is an entity which performs calculations over mental representations, or inner tokens or symbols which refer to features of the outer world. In short, the mind is thus viewed as a symbolic information-processing device, operating in a serial fashion, and governed by rules which are, at base, the language of thought.
Intractable problems confront GOFAI, according to Dreyfus, who recounts his major objections and lays the groundwork for a criticism as well for non-GOFAI modelling, specifically for Parallel Distributed Processing (PDP), or Connectionist, architectures and programming. To reinforce earlier GOFAI criticisms, he describes insuperable difficulties (outlined in 1972) confronting successful modelling of common-sense understanding, which requires a notion of relevance, contextually and holistically characterized, resulting from worldly, bodily experiences, not compatible with atomistic, symbolic data structures and discrete computations. He then develops the 1979-edition criticism which cites a further insuperable GOFAI problem: that of modelling the know-how requisite to judge relevance. 'Know-how,' or that activity of generalizing and determining relevance in an open-textured world constantly confronted, is argued to be not a matter of manipulating data, no matter how much is provided. Moreover, with serial processing strictures, symbolic computational attempts appear to be biologically unrealizable as well, and demand, in principle, procedures that require a combinatorial explosion of information that cannot be resolved by computational means alone.
More recent developments in AI change the landscape, for some dramatically. With the fashionable PCD, or Connectionist, models claiming to have overcome combinatorial bottleneck which plagued all von Neumannesque attempts, and with these non-serial and seemingly non-computational symbol-processing approaches to understanding the nature of the mental, Dreyfus wonders why anyone would cling to GOFAI. Besides, Connectionism even looks biologically correct, since brain design appears to be more like a pandemonium of parallel-processors, seeking some sort of neuronal equilibrium, than a central brain- in-a-brain operation carried out by a nervous system CPU surrogate. Why cling to GOFAI, indeed? And can, Dreyfus asks, the PDP architectures fare any better than GOFAI in describing the nature of human intelligence? Curiously, responses to both questions receive a similar treatment from him.
Apparently, GOFAI-lovers just don't seem to get it when it comes to appreciating what Dreyfus calls the Commonsense Knowledge Problem (CKP), alluded to in his earlier critcisms, nor---and this is new---do PDP models look to be any more promising. (This turns out to be a major criticism, and deserves more space than I can give it.) Basically, CKP is defined by three problems: (1) How everyday knowledge must be organized so that one can make inferences from it; (2) How skills or knowledge can be represented as knowing-that; and (3) How relevant knowledge can be brought to bear in particular situations (xviii). In fact, one might treat all three problems as ones involving selection of relevance. For example, learning to generalize is critical for intelligent behavior, but this requires associating inputs of the same type with successful decisions and actions. But in what does the relevant type consist? Relevance in this regard and in the context of ignoring and attending to features of novel settings as we confront them are not inherent in the context data, Dreyfus argues, but are, instead, relative to current situations in light of a myriad of human background experiences. What is relevant in one setting might not be so in another. Thus, whether by means of symbolic tokens (GOFAI), or having been learned through adjusted network connections (PDP), relevance is not to be gleaned by means of providing more information for the system to work with and through. Information about what is relevant only leads to circularity or a vicious regress of what is relevant to relevance for relevance for..... Admittedly, in narrowly-defined problem domains a machine might seem to pull off generalization skills, but this would be to mimic intelligence, at best, artificially enforced, as it were. Solving the CKP is a gauntlet Dreyfus lays down. Can the PDP paradigm solve it?
The philosopher Daniel Dennett, in Consciousness Explained (1991), thinks Dreyfus might very well believe so, at least indirectly, for he writes of two AI skepics (Dreyfus and John Searle) "Dreyfus has pledged allegiance to connectionism" (p. 270), referring to a 1988 piece written by Dreyfus and his brother Stuart Dreyfus (1988). Yet Hubert will shortly write in the current Introduction: "Indeed, neural-network researchers with their occasional ad hoc success but no principled way to generalize seem to be at a stage of GOFAI researchers when I wrote about them in the 1960's. It looks likely that the neglected and then revived connectionist approach is merely getting its deserved chance to fail" (xxxviii). With nothing short of harshness, the new anti-AI forces seem to draw yet another line in the sand.
For many AI system-designers, as well as for many scientifically- minded philosophers (myself included), Dreyfus' repeated 'It can't be done' stance will be unsettling, especially in light of his overt willingness to leave the putative mystery of the mind as simply that: a mystery. Yet, in fairness, he does raise serious objections that deserve and challenge equally serious counterarguments, presented carefully. Still, for those who do see a physicalistic basis of mentality as a correct and natural one, one should appreciate, in response to Dreyfus' skepticism, that machine models of human cognition need to square with neuroscience realities, just as brain mechanism models need to square with AI models that try to replicate intelligence. Combinations of multi- disciplinary approaches will no doubt emerge, with an aim to reproducing a system that actually learns to act intelligently in a world not circumscribed a priori by information overload or researchers' semantic paternalism. But we haven't heard the last of Dreyfus; maybe the 2001 revision will be titled What Computers and the Brain Can't Do (Either).

References:
(1) Dennett, Daniel C. 1991. Consciousness Explained . Little, Brown and Company: Boston, Toronto, London.
(2) Dreyfus, H. L., and Dreyfus, S. E. 1988. 'Making a Mind Versus Modeling the Brain: Artificial Intelligence Back at a Branchpoint,' in Graubard, S. R. 1988. The Artificial Intelligence Debate: False Starts, Real Foundations. The M.I.T. Press: Cambridge , Massachusetts.
(3) Haugeland, J. 1985. Artificial Intelligence: the Very Idea. The M.I.T. Press: Cambridge, Massachusetts.