Heidegger and AI

Why Heideggerian AI Failed and How Fixing it Would Require Making it More Heideggerian

This is a really interesting paper. In it, Hubert Dreyfus, known for his books What Computers Can’t Do, goes over why some of the more well-known AI projects have failed and also explores some worthwhile avenues where AI can succeed.

[In the 1960s] AI researchers were hard at work turning rationalist philosophy into a research program.

Dreyfus is referring to the Physical Symbols Theory of Newell and Simon that strove to empirically show that what is “really going on” in minds is the shuffling of symbols in a systematic way. By setting up the framework of AI in terms of this input>>processing>>output “boxology”, AI researchers attempted to demonstrate that the brain is really a very complicated information processor that could in principle be replicated on a silicon medium. After all, if all that matters is the “function” of information processing, then the actual substrate of the mind is irrelevant. All that matters is the algorithms, or “software”, running over-top the “hardware”. Notice that the entire research paradigm of AI, derived from cognitive science, is based on the metaphor of the computer. It is this metaphor that Dreyfus wants to combat and instead replace it with a more phenomenologically accurate account of what goes on when humans with minds interact with the environment.

Dreyfus uses the “frame problem” as a prime example of why this traditional symbol-shunting, representationalist program was doomed from the beginning. The frame problem is simply the problem of knowing the relevant context for a particular problem. AI programs need to know what particular knowledge is relevant to the the situation in order to realistically cope with the world. As Dreyfus is apt to point out, the human world of meaning is saturated with significance precisely because we are immersed in a “referential totality”. So for example, modeling the human use of tools can’t be done with “brute force” because whenever we use a hammer, the referential totality of nails and what-we-are-hammering-for comes into use. There is a particular way of being of hammers because they are embedded in a cultural “existential matrix” that is imparted onto the human world through the communal use of language.

Dreyfus concludes that in order for an AI to get past this crucial problem of contextual relevance, they would need to be imbued with particular “bodily needs” in order so that the AI could “cope” with the world. In other words, these AI need to be embodied and embedded in the world so that there is a particular significance for the program, or else it will never be able to act intelligently in the world. You can’t develop a truly artificial intelligence based on pure symbol shunting because the significance of the world stems not from our brain “processing” symbolically, but rather from the entire referential totality of culture. We can’t escape from the fact that our intelligence results from persons coping with an environment.

The Turing Test

In 1950, Alan Turing published a landmark paper in the journal Mind entitled “Computing Machinery and Intelligence”. In this paper he asked the question “Can machines think?” and proposed a method for determining whether a machine thought intelligently or not. This method became known as the Turing Test.

The test runs as follows, from wikipedia:

a human judge engages in a natural language conversation with one human and one machine, each of which try to appear human; if the judge cannot reliably tell which is which, then the machine is said to pass the test. In order to keep the test setting simple and universal (to explicitly test the linguistic capability of the machine instead of its ability to render words into audio), the conversation is usually limited to a text-only channel.

It is interesting to note that Turing himself thought that question itself(can machines think?) was “too meaningless to deserve discussion”. By this he meant that the most common objections to the question were usually drenched in emotional overtones to such a degree as to make them irrelevant. Nevertheless, Turing went on to discuss several objections to the idea can machines could ever properly be said to “think”.

Some of the objections he dismissed outright as ridiculous(such as the “head in the sand” objection that it would simply be too dreadful if machines thought), but others he gave more careful consideration of. The objection that I would like to discuss in this post is the “Argument from Consciousness” which denies the validity of the Turing Test because “No mechanism could feel(and not merely artificially signal, an easy contrivance) pleasure at its successes, grief when its valves fuse, be warmed by flattery, be made miserable by its mistakes, be charmed by sex, be angry or depressed when it cannot get what it wants.”

Turing counters this objection in the following way:

According to the most extreme form of this view the only way by which one could be sure that a machine thinks is to be the machine and to feel oneself thinking. One could then describe these feelings to the world, but of course no one would be justified in taking any notice. Likewise according to this view the only way to know that a man thinks is to be that particular man. It is in fact the solipsist point of view.

Turing goes on to note that if you do not accept this extreme viewpoint, then necessarily one must accept the validity of the terms of his test, namely that if a machine could fool you through textual-typing that it was human, then by all intents and purposes, that machine could be said to have thoughts. This might be seem silly at first because you might object and say “well, I can imagine a machine who thinks, but doesn’t have any emotions. It doesn’t care about anything”. The Turing test gets around this obvious objection because it postulates that what matters about minds is whether or not they can act in an intelligent way. He further argues that if any machine could act(type) in such a way as to convince any human observer that it was intelligent, then surely, it simply is intelligent. Furthermore, an example of intelligent thinking that would necessarily include an understanding of emotional overtones would be the reading of a good novel. This example illustrates the fact that emotion and intelligence are interlinked in such a way as to make it impossible to extract the two.

One might still object by saying that a machine could only possibly “represent” intelligent thoughts, but representations are not the same thing as real thoughts. My favorite philosopher Daniel Dennett has a fascinating reply to this objection. He asks us to imagine a computer simulation of a mathematician. Would it not be silly to complain that this simulated mathematician only gave mere representations of mathematical proofs, but not real proofs? Dennett, of course, says that representations of proofs are proofs because if this simulation of a mathematician produced proofs, would it not be valuable as a “colleague” to any proof-producing math department?

The moral of this simulated mathematician is that the criteria for what we consider thoughts depends not on whether it is represented or not represented, but rather, on the organizational pattern. In the same way that we would not care if a mathematical proof is “merely represented” if it is in fact a real proof, the question of whether “represented thoughts” were really thoughts becomes moot. We must take the Zen approach, and “unask” the question because it only obfuscates the important qualities of thought, namely its real-world effects.

Minds and machines

In this post, I want to briefly overview MIT’s exciting Cog project

Simply stated, Cog is “a set of sensors and actuators which tries to approximate the sensory and motor dynamics of a human body.” So what is the point of trying to replicate the “sensory and motor dynamics” of humans? Basically, the Cog researchers are trying to create an Artificial Intelligence.

In order to understand why an AI seemingly must have a humanoid body in order to be intelligent, one must have a basic understanding of embodiment theory.

The main thesis behind embodiment theory can be found in Shaun Gallagher’s How the Body Shapes the Mind. In this seminal work, Gallagher precisely defines the vocabulary necessary to talk about the thesis stated in the title: how the body shapes and influences the mind. Another overview of the embodiment thesis can be found here, by important embodiment researcher Andy Clark.

So, what does all this have to do with Cog and artificial intelligence? The MIT webpage has a nice overview and states “If we are to build a robot with human like intelligence then it must have a human like body in order to be able to develop similar sorts of representations.” Thus, the morphological(form) as well as the functional characteristics of our body-brain system play a critical role in shaping the dynamics of intelligent human interaction with the environment. The Cog project is not trying to “simulate” human intelligence on a symbolic level,which has been the traditional approach of Good Old-Fashioned Artificial Intelligence(GOFAI) but rather, is attempting to get human-level cognition to emerge from an intermodal and dynamic interaction with the environment.

Justification for the importance of Cog’s humanoid facial features is the fact that social interaction is perhaps the most important facet of human-environment-reciprocity that makes human intelligence uniquely human relative to the other great apes. It is the early prenatal and postnatal social-learning and development that gives rises to important conceptual constructs such as relativity(self/other, inside/outside, etc). If you are interested a neurological discussion of how such concepts arise from our embodiment, see my paper Mirror Neurons and the self

If this brief discussion of Cog as piqued your interest, you will probably be interested in some of MIT’s video overviews.

Lastly, I will end this post with another quote from the MIT page:

In any case….it turns out to be easier to build real robots than to simulate complex intereactions with the world, including perception and motor control. Leaving those things out would deprive us of key insights into the nature of human intelligence.