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?
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.
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.
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?
The the Alpha Go Discussion Post.
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.
If I understood correctly, this is only the first stage in the training of the policy network. Then (quoting from Nature):
Except that they don’t seem to use the resulting network in actual play; the only use is for deriving their state-evaluation network.