Is Google closer to creating a General AI than anyone imagined possible at this point in time? The possibility must be considered.
Originally shared by John Newman
Is Alphabet/Google’s AlphaGo an AGI?
AlphaGo’s architecture seems suspiciously human-like.
Even on Google’s announcement post, they state: “AlphaGo’s search algorithm is much more human-like than previous approaches […] During each simulated game, the policy network suggests intelligent moves to play, while the value network astutely evaluates the position that is reached.”
It goes on to say: “However, the most significant aspect of all this for us is that AlphaGo isn’t just an ‘expert’ system built with hand-crafted rules, but instead uses general machine learning techniques to allow it to improve itself, just by watching and playing games. […] Because the methods we have used are general purpose, our hope is that one day they could be extended to help us address some of society’s toughest and most pressing problems, from climate modelling to complex disease analysis.”
Is Google insinuating that this AI is not a “narrow” AI but a “general” AI, otherwise known as an AGI? Is that not a strong claim?
The paper is behind a paywall so I can’t say for sure [update: the paper is freely available. See comments below.] but the whole AlphaGo architecture looks suspiciously human-like to me. More specifically, this looks a lot like a mammalian brain. The policy network looks like a neocortex – it gets trained up by the inputs and the feedback from the value network. The value network is just like the animal’s saliency network. Our saliency networks feed information from our neocortex to our attention circuits in our frontostriatal loops, which in turn provide “top down” signals to the neocortex.
Just as human saliency networks feel the results of neocortical computation, (but do not experience them directly) this value network feels the inputs from the policy network and then feeds back new salience signals to the policy network, coloring that input with an emotional context from the situation. This way it develops an opinion of those inputs. This is starting to smell like animal cognition to me.
So, has anyone had a chance to read it? When they say this algorithm is more human-like, how literally do they mean that?