4 ms·

I hope the next focus for Deepmind will be on opening up the black box and trying to put some explainability aspects in place, as has been done before for other

I hope the next focus for Deepmind will be on opening up the black box and trying to put some explainability aspects in place, as has been done before for other deep learning architectures (or at least, people are starting to play around with). So many questions of the form "I wish we could ask AlphaGo what is was thinking here" -- so this would be great to have.

orthoganol 9y ago
Probably something like "I've been trained on many similar past situations, these are possible moves that worked out well, let's simulate a few of them internally, OK, here's my move." Full disclosure, I really don't know much about the internals of DeepMind, but if it's just a DL system on steroids with sampling, there isn't really any 'thinking' happening, it's just tapping probability distros over possible moves conditioned on tons of training data.
- Macuyiko 9y ago
  I didn't downvote you, but there's more going on here than just inferencing over a database of prior knowledge. The game is too complex for that. Compare this approach for instance to Monte Carlo based approaches which aren't doing so well. Speaking of Monte Carlo, the Lee Sedol version of AlphaGo did combine a series of deep networks with Monte Carlo sampling, but rumor is that was replaced in this version altogether (it's also running on TPU's now). Would like to see some more technical details as well.
  - computerex 9y ago
    Their state of the art approach hybridizes MCTS with deep neural networks. I'd be interesting indeed if they manage to get better performance without MCTS all together, although they did achieve very impressive performance without MCTS.
- taneq 9y ago
  How do you define 'thinking' such that an artificial neural network can't do it, but a biological neural network can?
  - orthoganol 9y ago
    Am I doing that? I don't think "I wonder how a sea-turtle thinks" is necessarily on a different plane than how our machines today 'think', on a high level. Or by biological NN did you only mean humans? If you only meant reasoning humans, it strikes me as pretty obvious they are different. Didn't Andrew Ng define the scope of DL as anything a human can do in less than a second, only? I don't think the "neural" analogy is meaningful or interesting anyways, and I don't think DL people do either. Layers of logistic regression units isn't a brain, or would you argue brains are logistic regression layers?

orthoganol 9y ago

Probably something like "I've been trained on many similar past situations, these are possible moves that worked out well, let's simulate a few of them internally, OK, here's my move." Full disclosure, I really don't know much about the internals of DeepMind, but if it's just a DL system on steroids with sampling, there isn't really any 'thinking' happening, it's just tapping probability distros over possible moves conditioned on tons of training data.

Macuyiko 9y ago
I didn't downvote you, but there's more going on here than just inferencing over a database of prior knowledge. The game is too complex for that. Compare this approach for instance to Monte Carlo based approaches which aren't doing so well. Speaking of Monte Carlo, the Lee Sedol version of AlphaGo did combine a series of deep networks with Monte Carlo sampling, but rumor is that was replaced in this version altogether (it's also running on TPU's now). Would like to see some more technical details as well.
- computerex 9y ago
  Their state of the art approach hybridizes MCTS with deep neural networks. I'd be interesting indeed if they manage to get better performance without MCTS all together, although they did achieve very impressive performance without MCTS.
taneq 9y ago
How do you define 'thinking' such that an artificial neural network can't do it, but a biological neural network can?
- orthoganol 9y ago
  Am I doing that? I don't think "I wonder how a sea-turtle thinks" is necessarily on a different plane than how our machines today 'think', on a high level. Or by biological NN did you only mean humans? If you only meant reasoning humans, it strikes me as pretty obvious they are different. Didn't Andrew Ng define the scope of DL as anything a human can do in less than a second, only? I don't think the "neural" analogy is meaningful or interesting anyways, and I don't think DL people do either. Layers of logistic regression units isn't a brain, or would you argue brains are logistic regression layers?