Roberto Cordeschi

Philosophical Assumptions in Artificial Intelligence: A Tentative Criticism of a Criticism

It is well known that, as a result of the development of computability theory and computer science, certain higher processes usually called “intelligent” have been proved as algorithmic, contrary to previous assumptions, Thus the hypothesis arose that it was possible to find a new level of explanation which could be crucial for the study of mind. In the course of the short history of AI, the hypothesis became more explicit. More recently Allen Newell (1980) discussed it as the physical symbol system hypothesis, widely accepted both in (classical, not connectionistic) AI and in cognitive science.The level of explanation chosen as the crucial one was the symbol level.

New Steps Towards the Artificial

While denying the relevance of cybernetic robotics, Newell and Simon pointed out that, besides symbolic information-processing modeling, another, different line of research was being pursued at the time, which would use formal models that simulated the properties of neurons and of their organization into nets (Newell and Simon, 1963: 392). In fact, at the time of writing, this line of research was well established, and some of the events in its development were touched on in previous chapters. Rashevsky had tried, at the end of the 1920s and later in his 1938 Mathematical Biophysics, to analyze various physiological and neural phenomena mathematically. In 1943, McCulloch and Pitts introduced Boolean algebra to describe nets consisting of formal neurons. In 1949, in The Organization of Behavior, Hebb formulated his law on the modification of nerve connections during learning. Nets of formal neurons, with connections that could be modified through rules of this kind, have been simulated on computers since the early 1950s.

Behavior, Purpose and Teleology

We have mentioned the paradox of backward causation, which afflicted the explanation of purposeful behavior. This kind of behavior seemed to consist in an inversion of the causal sequence: the future or final result of an action (the end to which it is directed) precedes or anticipates the action. “A sequence entirely inadmissible in a world of pure mechanism,” was how McDougall (1923a: 194) saw it. What was known as the metaphysical solution of the paradox always saw the end or the goal of an action as its ‘fmal cause,’ and teleological explanation was considered completely different from causal explanation.1 Thus an unbridgeable gulf opened up between teleology and science, one that challenged the very plausibility of using teleological language in the scientific study of the behavior of living organisms.

The Many Forms of the Artificial

The public emergence of cybernetics can be dated rather sharply because of the nearly simultaneous appearance [in 1943] of three basic papers on the subject. Rosenblueth, Wiener and Bigelow showed how simple goals and purposes could be realized in feedback machines. McCulloch and Pitts pointed out how some other logical categories and mental concepts could be represented in ‘neural’ nets, and Craik suggested a variety of ways by which machines might use models and analogies. To be sure, all of these had their own intellectual ancestors, but here for the first time we see a sufficiently concrete (i.e. technical) foundation for the use of mentalistic language as a constructive and powerful tool for describing machines (M. Minsky).

The Robot Approach

More than once we have had occasion to note that a form of learning which might be described in terms of Pavlov’s conditioned reflex was quite familiar to modern psychological and neurological associationism. When the West, especially the United States, learned of Pavlov’s work, psychologists posed the question of how conditioning might fit into the framework of the different forms of learning that had already been studied.1 In Watson’s version, conditioning, understood as stimulus substitution, could be explained by the temporal contiguity of the conditioned stimulus to the unconditioned stimulus and to the response to be conditioned, i.e. the response that was evoked initially by the unconditioned stimulus. Watson pointed to temporal contiguity of stimulus and response to explain other forms of learning as well, including selective or trial-and-error learning, which according to Thomdike was based on the law of effect. Thomdike did not ascribe much importance to conditioning as stimulus substitution (or “associative shifting,” as he put it). He considered conditioning a very limited form in comparison with selective learning and, in any case, a form based on other principles. From his point of view, if learning by conditioning could result from simple contiguity, more complex forms such as selective learning could be explained only by the strengthening of certain S-R connections as set forth by the law of effect.

Chemical Machines and Action Systems

Jacques Loeb’s investigations into the behavior of numerous lower organisms are justly famous. In particular, he carefully documented the way in which the orientation of bilaterally symmetrical lower animals depended on light, i.e. phototropism, a phenomenon that the botanist Julius Sachs had studied in plants. And Loeb’s systematic monograph, Comparative Physiology of the Brain and Comparative Psychology published at the dawn of the twentieth century (Loeb, 1900), made his work popular with researchers in various fields.

Cybernetics and the Origins of Artificial Intelligence

In the previous chapters we identified various stages in the discovery of a novel strategy in the study of mind and behavior, suggested by ‘new’ notions of the machine. More than once we alluded to cybernetics, especially in Chapter 4. We mentioned there claims made by Rosenblueth, Wiener and Bigelow in their 1943 paper, emphasizing the role that began to be attributed to feedback in the study of various aspects of organism behavior. That paper is usually considered the manifesto of what Wiener called cybernetics, i.e. the study of “control and communication in the animal and the machine” (Wiener, 1948/1961).

Nervous Conduction: Early Connectionism, Physical Analogies and Machines

In the previous chapter we saw that for Loeb associative memory was a mark of intelligence and learning capacity. His thesis was that the lower organisms did not have associative memory, a thesis that seemed to be corroborated by a machine that was actually built by an engineer, Benjamin Miessner. The machine displayed behavior that was indistinguishable from that of a lower organism, and since it was a machine, and since it was automatic, it was by definition without associative memory or the ability to learn.