Re:Philosophy Has Made Plenty of Progress
by visarga on Wednesday November 07, 2018 @11:32AM
(#18400214)
The notions of agent, environment and reward are encountered in Reinforcement Learning which is a sub-field of ML but also relevant to biological agents. There is a great RL course on YouTube by David Silver (one of the creators of AlphaGo, which is probably the most famous application of RL so far).
The course can properly set the perspective of RL and shed new light on the philosophy of mind, if you take what it says and then extrapolate to humans and other agents.
What I find fascinating about RL is that it can be defined concretely. Consciousness can only be defined by reference to other words, in a less exact and concrete way. RL can also explain how meaning appears, based on future reward prediction. The rich sensations we have can be explained by an encoding-decoding architecture based on reconstruction error. Many difficulties in RL map back to difficulties humans have in choosing how to act - the exploration-exploitation tradeoff, instinct vs reason (two different ways to perform RL, one based purely on rewards and the other based on an internal model of the world and rewards). Some of the problems related to multi-agent RL are also covered in Game Theory, such as the prisoner's dilemma.
Regarding philosophy of language: we have today numerical representations of meaning in words and phrases. They are usually represented as vectors in high dimensional space or sets of vectors. For example it is customary to use 300-1000 dimensions for representing words. These vectors have a nice property - the closer two words are in meaning, the smaller the angle between the two vectors. They are derived by trying to predict a word by the context, or vice versa, on a corpus of text several Gb in size.
Many ideas from the philosophy of language, such as the meaning of words being related to the 'game' being played (activity with a purpose), emerge naturally from successful AI models. I'd say that where philosophy had a glimpse, AI has a testable implementation that can solve real world problems. Where philosophy uses mere words, AI uses probability distributions and datasets to define such models. The brain is probably doing something similar.
Other things that AI has managed to do so far: to encode images into latent representations and back, to synthesise images. Same for speech - we have speech recognition and text to speech. Some modules used in neural networks are analogous to imagination, attention, memory, emotion, intuition and many other aspects of the mind. On narrow domains computers can already best humans at perception.
The piece that is missing in AI to match human level is the prior knowledge encoded in our genes. We have been optimised to learn and function well in our environment by evolution. That means our verbal areas in the brain have a notion of invariance to time translation, visual areas have invariance to space translation, and conceptual areas have an invariance to permutation. There might be more invariances but we just don't know yet and that's why AI models are not yet up to par with humans. But we can still learn a lot about ourselves by analogy with AI agents. And that's where I think philosophy should listen.
https://www.youtube.com/watch?v=2pWv7GOvuf0
The course can properly set the perspective of RL and shed new light on the philosophy of mind, if you take what it says and then extrapolate to humans and other agents.
What I find fascinating about RL is that it can be defined concretely. Consciousness can only be defined by reference to other words, in a less exact and concrete way. RL can also explain how meaning appears, based on future reward prediction. The rich sensations we have can be explained by an encoding-decoding architecture based on reconstruction error. Many difficulties in RL map back to difficulties humans have in choosing how to act - the exploration-exploitation tradeoff, instinct vs reason (two different ways to perform RL, one based purely on rewards and the other based on an internal model of the world and rewards). Some of the problems related to multi-agent RL are also covered in Game Theory, such as the prisoner's dilemma.
Regarding philosophy of language: we have today numerical representations of meaning in words and phrases. They are usually represented as vectors in high dimensional space or sets of vectors. For example it is customary to use 300-1000 dimensions for representing words. These vectors have a nice property - the closer two words are in meaning, the smaller the angle between the two vectors. They are derived by trying to predict a word by the context, or vice versa, on a corpus of text several Gb in size.
Many ideas from the philosophy of language, such as the meaning of words being related to the 'game' being played (activity with a purpose), emerge naturally from successful AI models. I'd say that where philosophy had a glimpse, AI has a testable implementation that can solve real world problems. Where philosophy uses mere words, AI uses probability distributions and datasets to define such models. The brain is probably doing something similar.
Other things that AI has managed to do so far: to encode images into latent representations and back, to synthesise images. Same for speech - we have speech recognition and text to speech. Some modules used in neural networks are analogous to imagination, attention, memory, emotion, intuition and many other aspects of the mind. On narrow domains computers can already best humans at perception.
The piece that is missing in AI to match human level is the prior knowledge encoded in our genes. We have been optimised to learn and function well in our environment by evolution. That means our verbal areas in the brain have a notion of invariance to time translation, visual areas have invariance to space translation, and conceptual areas have an invariance to permutation. There might be more invariances but we just don't know yet and that's why AI models are not yet up to par with humans. But we can still learn a lot about ourselves by analogy with AI agents. And that's where I think philosophy should listen.