12 hours ago, happiness said:
Today I saw a demonstration of an app that counts objects of the same type in a picture. The software isolates a similar characteristic among concretes in a visual field and performs an act of induction to form a concept. Computers are using human epistemology. Others have already noted the similarities between Objectivist epistemology and object oriented programming.
Is it possible that a sufficiently advanced form of AI, obeying a rational epistemology and lacking the capacity for evasion, tell us that Objectivism is the correct philosophy, and that capitalism is the correct social system?
I suggest that a machine—say, a learning machine such as an artificial neural network one—has not educed a human concept even if it has been designed to learn dimensions of similarity among a group of items and even if its groupings according to degrees of similarity along those dimensions are registered by measure values and even if a label for each of those comparatively similar-member collections were to be given by the machine, these distinguished collections would not be like human concepts, and for three reasons:
(1) Human perceptual comparative-similarity groupings are made against a background of possible actions upon them and uses for them by the agent who is on his way to forming a concept. This is contemporary Ecological Psychology continuing its research down from James and Eleanor Gibson, who acknowledged that their leading idea of “affordances” in perception had been a gift from William James and John Dewey. Rand, Peikoff, and Kelley did not put enough emphasis on this aspect of perception. Rand did set out that while the human is learning what things are, he has a parallel assessment going on as to whether the item might be to be avoided or might be desirable. Rand once mentioned, correctly, that most concepts are amenable to definition. In my 1990 paper “Capturing Concepts” I proposed than prior to learning to make sentences, the toddler (all of us) embed our single-word utterances and concepts into action-schemata. To get nearer to human concepts, even the most elementary concepts, a machine probably would need to be a robot, an agent, given a set of values and their interrelations by human designers and given ability to register and assess affordances. Perhaps the lab at MIT has been working on this.
(2) Human perceptual learning is as part of a process of development towards acquisition of discursive thought and communication. Single-word stage of human conceptual consciousness and the predicative multi-word stage are motivated very much from urge to more and more precise communication with other humans. With this motivation not attending machine learning, and coloring its concepts and their interconnections, I think machine concepts would be but a stick-man of ours. Indeed, getting outputs from the learning machine we desire does make the machine operations in some community with humans, though not directly with other learning machines. This condition and its profundity in human conceptualizing was silently passed over by Rand, but it should not be neglected in a fully realistic picture of human conceptual operations.
(3) A machine able to learn comparative similarity groupings among items would be doing something that humans can do, though perhaps without the affordances and background sociality of human cognition concerning the items. Other analyses of similarity computations besides the measurement ones given by Rand have been set out in the psychological literature. One could program a machine to detect particular similarities using these various computational schemes, but unless the results have different advantages, I don’t see how one could determine whether Rand’s measurement-analysis of similarity was receiving some confirmation that it is the better. And in the case of learning machines, I’m unsure if it can be determined which of the computational schemes is doing the work in learning to sort. Further, to show such sorting capability does not show conceptual ability. If a test for conceptual ability could be shown—say, passing a Turing test—and it were shown that machines using Rand’s measurement-omission scheme for forming concepts from similarity groupings were the most successful in the machine, then we might say Rand’s distinctive idea concerning the nature of concepts has received some recommendation from trials in machines. But that is a big IF, and unless we take passing a Turing test as showing understanding (and using sets in knowing concepts and numbers!), we’d not want to conclude that the machine has human-like concepts at all. And between you and me and the fence post, I don’t think understanding at all is possible without the agent being conscious and, therefore, alive.