Hacker News Reader

Comment by paraschopra | Hacker News Reader

For anyone who is interested in a tutorial and actual implementation of active inference (an idea based on Free Energy Principle), here's one in Python https://kaiu.me/2017/07/11/introducing-the-deep-active-infer...

I have been trying to understand FEP, and so far my understanding is that essentially the agent tries to learn the generative model that most closely explains observations and then tries to act in a way that are more likely to cause the environment to generate its preferred observations (say pH and temperature in the right range).

The problem with this approach is in scalability of inference and candidate model generation. By the time you provide model for the agent, you as a designer have coded much of your knowledge already and hence constrain the agent. True AI will build model from the scratch, and not just learn model complexity.

View on HN

>True AI will build model from the scratch, and not just learn model complexity.

There's no such thing as truly learning "from scratch" -- the No Free Lunch Theorem holds no matter what. What you can do is find a sufficiently large (ex: Turing-complete) hypothesis class, and make simplifying assumptions to allow it to be feasibly learnable (such as regularization or priors).

The No Free Lunch Theorem is irrelevant to the real world [0][1]. It assumes all functions, even those with infinite algorithmic complexity, are equally likely.

You should look into algorithmic probability for a better foundation.

[0] http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.540....

[1] https://arxiv.org/abs/1111.3846.pdf

On the one hand, yes, the No Free Lunch Theorem seems to intuitively rely on the set-theoretic definition of functions, rather than building on a constructive foundation to hypothesize that functions which are "physically harder", in some sense, are less likely.

On the other hand, algorithmic probability requires first defining a Turing machine, rendering the Solomonoff Measure defined only up to a specific programming language, which can bias it some arbitrary amount. That's on top of the Solomonoff Measure itself being incomputable, and so utterly useless as a foundation for real-world machine learning and computational cognitive science.

I agree that positing a Bayesian prior on functions/programs/causal structures gets you around the No Free Lunch Theorem. The question just then ends up being: what sort of hypothesis space, and what sort of prior, sufficiently resemble the real world (the data-generating process) to allow for learning from a given data set? That's a matter of science.

I'm not sure I follow you.

I didn't mean intelligence in any abstract problem-space. I meant intelligence in the world and type of problems we humans deal with (in fact, I'm unsure what process should we call intelligence in non-human context).

In the context I'm talking about, we at least have one algorithm that has build models from the scratch: evolution by natural selection.

The idea is that you learn a model by calculating the derivative of free energy with respect to your model parameters.

Yes, but you have to specify a generative model (or at least put boundaries to it). Then you learn params of that model.

I was talking about learning the model structure also.

Some attempts have been made in the form of Bayesian model reduction [1].

The idea is to 'carve' out the structure of your model using free energy minimization.

[1] https://arxiv.org/abs/1805.07092

Reply on news.ycombinator.com