[satvikpendem]

I am not sure if this is covered in the podcast, but from the title, I wonder if in deep neural net research there are papers and implementations of a system 1 and 2 like model. One part of the neural net could be system 1 for fast decision making, and other could be system 2, with more deliberate analysis. They could communicate over some shared medium.

[K0SM0S]

I got your answer! He indeed explicitely says in the interview that deep learning is really system 1 only. That was surprising but makes total sense — it's an unconscious, automatic, fast, effort-free response, which is exactly what system 1 is.

Note that both system 1 and 2 are trainable and able to perform complex tasks (he takes the example of a chess player for whom only strong moves come to mind, which is system 1 playing chess in that case; and system 2 is more of a "validation" for system 1's output in such a case).

He doesn't go into it but I think you could make the reverse argument, that system 2 is "checked" by system 1, when we "feel" that something, even though "correct", is just "not right" for instance. That kind of judgement over a thought or idea is nearly instant, it's very system-1 like, and keeps popping up in our thinking as we "judge" said thoughts and do some triage as we go along.

As for AI and system 2, the problem is that system 2 is conscious, deliberate, and aware of causality and meaning — and the last two are really hard problems for now. He mentions earlier ML models pre-DL (when they tried to do it the hard, symbolic way iirc?), and indeed the question of whether current architecture can or cannot generalize up to system 2 is opened. Yann Lecun apparently thinks it can (just that we don't know if it's right around the corner or very, very far away), Lex (and most AI experts I heard) think not, that there's a fundamentally 'other' kind of architecture(s) required.

[laughingman2]

Would you consider using neural nets to solve integration and differential equations as it doing system 2 reasoning?

https://ai.facebook.com/blog/using-neural-networks-to-solve-...

Or what tasks are in the domain of system 2?

[K0SM0S]

Sorry, late reply, hope you get this!

I believe it's totally "system 1", and actually by design.

First of all it's not a new kind of NN, it's more about applying a given problem to another technique, name consider math syntax as just another kind of language:

> represent complex mathematical expressions as a kind of language and then treating solutions as a translation problem

(which might seem obvious but I guess it took that much refinement to yield actual results)

Now, take these quotes, emphasis mine:

> Humans who are particularly good at symbolic math often rely on a kind of intuition. They have a sense of what the solution to a given problem should look like

> By training a model to detect patterns in symbolic equations, we believed that a neural network could piece together the clues that led to their solutions, roughly similar to a human’s intuition-based approach to complex problems.

Intuition, intuition-based approach: this is exactly what system 1 represents.

Also note the results:

> Our model demonstrated 99.7 percent accuracy when solving integration problems, and 94 percent and 81.2 percent accuracy, respectively, for first- and second-order differential equations.

One major difference between systems 1 and 2 is that 1 is fuzzy, intuitive, it's not always exact, it's very analog; whereas system 2 is able to be correct, exact, precise — and like the researchers themselves validating the 5,000 answers, you'd expect a "well trained" math intelligence to solve 100% (or close enough) of these problems. It may take time but give yourself 20 years and you'll get there no doubt; whereas this narrow language-AI with one hundred million examples still makes mistakes.

Very system 1 indeed.

[glial]

Typically the two systems are called 'model-free' and 'model-based' learning.

Model free (system 1) is fast, stimulus-response mappings. In ML, these mappings are called a policy. Most of reinforcement learning, including deep-learning-based RL, is model free.

Model-based (system 2) is less widely used. In this case, an agent or system is trying to learn the dynamics of a system and use those to project or forecast into the future. This is really helpful for e.g. learning a control system. Being able to use a model of the world to make accurate predictions lets you plan.

[p13rr0m]

I believe the paper SlowFast Networks for Video Recognition goes somewhat in that direction. The architecture is split into a fast pathway, to capture motion information, and a slow pathway that captures spatial semantics. https://arxiv.org/pdf/1812.03982.pdf