3

As a machine learning researcher and amateur philosopher, I'm interested in the current status of the debate between the symbolic AI and connectionism\pattern-recognition schools from the perspective of philosophers. From my experience and some basic reading I've done (see for example the connectionism Wikipedia page) it appears that while in neuroscience and the computer science AI community, the neural net (connectionist) based approaches are the mainstream, in psychology and philosophy there isn't a widespread acceptance of them.

I couldn't find any citations to either support or counter the last claim (regarding the lack of widespread acceptance in philosophy), which would maybe be implicitly supporting it.

Also, and this is a more open question so I hope it's still appropriate- what do you think some main implications of the connectionist model would be on philosophy? For example, to me it seems that it could lead to more emphasis on practice\education methods, context and experience, reliance on intuition and "soft definitions" (sounds somewhat Zen...), in contrast for example to emphasis on absolute notions, definitions and abstractions.

  • Neural networks provide a very important model for how learning takes place, but I disagree with your assumptions. Patterns are abstract notions, and if it weren't for the guidance of programmers, computers would never be able to recognize patterns on their own because they simply don't know what to look for. Neural network research is confirming what Kant said long ago about a priori knowledge being a necessary prerequisite for learning. Computers get their a priori knowledge from programmers. – user3017 Jun 26 '16 at 1:43
  • Actually, I think that neural network research is leading a trend where computers need less and less programmer supplied prior knowledge, who let the data + network architecture do more of the work instead of the a priori knowledge. One could argue that the data and architecture are themselves a form of prior knowledge and in a sense this is true, but the point is still that the emphasis is shifted more to letting the computer experience the data first-hand and learn for itself, and less to applying our preconceptions to it. – RonenKi Jun 26 '16 at 10:52
  • I don't disagree with that. Like I said, they provide an important model. However, a priori knowledge still appears to be an unavoidable fact. To demonstrate otherwise would involve feeding a computer data and leaving it up to the machine to realize that the data might be useful for something and that pattern recognition just might be the way to discover what it is. Kind of like hammers and architects. The importance of the hammer doesn't demonstrate that architects are not needed. – user3017 Jun 26 '16 at 11:18
  • @PédeLeão, I believe that you are totally wrong. There are may examples in which neural networks learn patterns without anyone telling them what to look for. For example the famous network that learned to play atari games - youtube.com/watch?v=Q70ulPJW3Gk. what for you is the video game screen, was a meaningless vector of numbers for the network. Its only objective was to maxime the final score. – nir Jun 26 '16 at 13:25
  • @nir. You're misunderstanding my point: 1. The neural network serves as a functional definition of a pattern, so the presupposition of the utility of patterns is programmed into it, rather than learned. 2. It doesn't know what to look for in a good pattern and has to be told by some sort of training or feedback. The network you're talking about wouldn't know a good pattern from a bad pattern without the scores. – user3017 Jun 26 '16 at 14:07
2

First of all, a clarification is necessary: Connectionism and Pattern Recognition are not the same thing. Neural Networks are one pattern recognition method among many, and in particular Random Forests and Support Vector Machines are very popular among machine learning practitioners without having anything to do with the way the human mind works. In fact, for a while in the mid to late 2000s, it was thought that RF and SVM had completely replaced NNets as the dominant machine learning paradigm, and it was only recently that Deep Learning NNets have made a (debatable) comeback.

Even when they were dominant, the most successful model - Multilayer Perceptrons trained by Backpropagation - had vey little in common with real neurons other than the general concept of representing learning problems as networks and nodes. So saying that NNets are mainstream among the AI community is a stretch, and insofar as they are popular, they are only so because they achieve results, not because they are close to the way the real human minds works.

Now for the answers: Connectionism appears in at least two topics of philosophy of mind.

  1. Connectionism provides an alternative to those who on one hand do not believe that the human mind is just a biological Turing machine (under the general heading of mind-body functionalism/computational theory of mind) but at the same time do not want to subscribe to any form of mind-body dualism. However, this is premised on the idea that neural networks are somehow capable of super-Turing computation, and this itself is a minority view among computer scientists. In particular for neural networks to be able to go beyond the Turing limit, they would have to be implemented on an analogue computer with unbounded precision, otherwise they face the same limitations that any Turing machine faces.
  2. Connectionism is considered an argument against the idea that the mind has an inherent language built into it form birth - see for example Noam Chomsky's universal grammar concept or Jerry Fodor's Language Of Thought Hypothesis. Theories such as Chomsky's and Fodor's are actually modern versions of the 17th century concept of innate ideas defended by DesCartes and other rationalists, that there is some knowledge (e.g. logic and reason, moral good, etc...) that we are born with beforehand. Opposed to the theory of innate ideas was Locke and other empiricists Tabula Rasa proposal, that the human mind is a blank slate upon which thoughts are gradually imprinted using sense data. Connectionism is an argument against innate idea type theories and in support of blank slate type theories of how the mind functions, and as such falls into the overall rationalism vs empiricism debate.

Keep in mind that the issue in (1) is an ontological/metaphysical question about what the mind is, whereas the issue in (2) is more of cognitive science psychology question of how the mind works, although both are related. In particular, it possible that the mind is ontologically a neural network, but still has some functional aspects that support the Fodor/Chomsky camp against pure empiricists.

Philosopher of mind David Chalmers has a good collection of references for connectionism with regards to questions of philosophy of mind.

  • Thanks for the comprehensive answer. There seems to me an important question alongside the question of expressive power of NNs, and that is the training process (optimization, data, etc..).These are the problems encountered by practitioners, for whom questions of theoretical limits aren't so relevant- NNs can be shown to be universal function approximators, but the problem is training them. It seems to me this may have reflections in philosophy- practical questions of human learning\training gaining importance alongside the theoretical limit questions (mind\Turing machine equivalence, etc..) – RonenKi Jun 29 '16 at 21:36
  • To clarify, obviously such questions of human learning have traditionally been handled in the framework of the philosophy of education. But I think neural networks have started a trend of theoreticians leaving the theoretical questions and moving more towards the practical ones (without the latter being seen as having lesser quality), and I wonder if there may be an equivalent shift in philosophy. These are just my impressions, hard to find citations for these sort of things, so just wondering if they resonate with others. – RonenKi Jun 29 '16 at 21:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.