New neural networks coming to PA

I understand... Nothing!

What I gathered: AI gets more cleverz. Yay!

So.... you took the day off of work so that you could work on PA?

I like you.

Really nice to see things like that are still being considered in the stage the game is! Also, finally a blog update !
Awesome!!

Fascinating! Thanks for sharing.

How much CPU time is the AI currently allocated? If indeed it is even capped?

Really great read, thank you for sharing Sorian!

Awesome!! Will this allow the AI to kill us even more efficient?

Appears reasonable so far, maybe except for that part where you bound the activation function to layer types rather than actual layers.

How do you scale the hidden layers? Same number of nodes for each layer, or configurable width?

input -> rectifier -> rectifier -> sigmoid -> output, or how does your current test setup look like?

Say hello to your new robot overlords.
#skynet

For the intellectually challenged among us (IE me), what tangible impact can we expect to see from your redesign of the neural networks?

The AI will be the next King of the Planet. You heard it here first!

Great stuff! I'm just starting in machine learning and network modelling, but I work at the institute where Demis Hassabis did his PhD and founded Deep Mind. I take it that it requires more than just multiple layers for this approach to qualify as so-called 'deep learning'? My basic understanding is that the recent movement towards deep hierarchical networks isn't fundamentally different from that of previous neural net implementations, the networks themselves are just more complex with more layers (and associated changes in algorithms to handle that complexity)? It's interesting to know that as you move upwards in the hierarchy learning becomes less sensitive to input (sidenote - my research suggests this sensory insulation is also true of the brain's deepest, most central networks!). As I understand it in the context of decoding, this allows the higher levels to encode gradually more complex features that integrate over multiple lower level inputs? E.g. in a neural net with enough layers one should expect the highest levels to code the most general categories. It makes sense that these should be insulated from the random noise of regular input; curious to hear more about how your changes will impact the interaction between high and low levels. Anyway i'm just curious how this plays out for PA's AI. Does the highest level encode something like meta-strategies? Could you improve the ability of the AI to respond to discrete events (i.e. snipes) by adding more layers? Also I won't ask you why the network isn't Bayesian as Demis recently revealed even approximate Bayesian inference is too costly for deep nets. But damn if it wouldn't be cool if PA's brain was Bayesian!

So cool when my work and play collide

edit: confusing as I just realized 'top' and 'bottom' and 'input' and 'output' are reversed in your parlance from the typical use in neuroscience (bottom = input). Also the sidebar on your blog is jittering intensely on chrome for me!

Oh my. Who does that? Everybody knows that information flows downward (always!).

Haha, basically all of neuroscience

Well, maybe not if you are really into the Bayesian brain, then you would at least view information processing as a reciprocal process. From the classical view however (otherwise known as 'feed-forward'), information flows 'upwards' or 'forwards' from the sensory epithelium to the cortex, is manipulated, and then transformed into an action signal (backwards/output). Really it's all a perception-action circle though!

This is why I studied computer science and never turned to neuroscience, so many weird ideas about information.
Now, I go back to my desk and draw more trees.

You rule anyway! If you want to pretend to be a (theoretical) neuroscientist, you could hook up the entire C Elegans connectome to PA and see if it is a better player than @elodea. I'm sure there is a nature paper in that

Sounds like some awesome work, I look forward to seeing that wiggle and pondering the network behind it!

Essentially, but also much more than that. Interestingly we've known the complete 'connectome' (map of the entire network and all connections) of the c elegans for nearly 50 years but it hasn't led to many breakthroughs. This group is trying to get the whole thing (it's only about 300 neurons in total) simulated online and open source so people can hook it up to funny things like worm-robots. In practice I doubt it would be very useful for your application (which I think is orders of magnitude more simple than even this network?) but it sure would be funny (and buzzworthy).

edit: the future of PA?