You're looking at Less Wrong's discussion board. This includes all posts, including those that haven't been promoted to the front page yet. For more information, see About Less Wrong.

bogus comments on Open thread, Jan. 25 - Jan. 31, 2016 - Less Wrong Discussion

3 Post author: username2 25 January 2016 09:07PM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Article Navigation

Comments (169)

Comments (169)

You are viewing a single comment's thread. Show more comments above.

Comment author: MrMind 28 January 2016 11:17:21AM *  0 points [-]

Those that follows are random spurts of ideas that emerged when thinking at AlphaGo. I make no claim of either validity, soundness or even sanity. But they are random interesting directions that are fun for me to investigate, and they might turn out interesting for you too:

  • AlphaGo uses two deep neural networks to prune the enormous search tree of a Go position, and it does so unsupervised.
  • Information geometry allows us to treat information theory as geometry.
  • Neural networks allows us to partition high-dimensional data.
  • Pruning a search tree is also strangely similar to dual intuitionistic logic.
  • Deep neural networks can thus apply a sort of paraconsistent probabilistic deduction.
  • Probabilistc self-reflection is possible.
  • Deep neural networks can operate a sort of paraconsistent probabilistic self-reflection?
Comment author: bogus 29 January 2016 10:25:15PM 0 points [-]

AlphaGo uses two deep neural networks to prune the enormous search tree of a Go position, and it does so unsupervised.

lol no. The pruning ('policy') network is entirely the result of supervised learning from human games. The other network is used to evaluate game states.

Your other ideas are more interesting, but they are not related to AlphaGo specifically, just deep neural networks.

Comment author: MrMind 01 February 2016 09:32:05AM 0 points [-]

lol no. The pruning ('policy') network is entirely the result of supervised learning from human games.

If I understood correctly, this is only the first stage in the training of the policy network. Then (quoting from Nature):

The second stage of the training pipeline aims at improving the policy network by policy gradient reinforcement learning (RL). The RL policy network pρ is identical in structure to the SL policy network, and its weights ρ are initialised to the same values, ρ = σ. We play games between the current policy network pρ and a randomly selected previous iteration of the policy network.

Comment author: bogus 01 February 2016 08:04:53PM 0 points [-]

The second stage of the training pipeline aims at improving the policy network by policy gradient reinforcement learning (RL).

Except that they don't seem to use the resulting network in actual play; the only use is for deriving their state-evaluation network.

Powered by Reddit Powered by Reddit