4 ms·

I would argue that there is actually an unrecognized inverse of this effect, too. People have a tendency to conflate domain-specific human-level performance wit

I would argue that there is actually an unrecognized inverse of this effect, too. People have a tendency to conflate domain-specific human-level performance with "intelligence" (for some definition of that word). Then suddenly we jump to "DeepMind won at go, therefore humanity is on the brink of creating Skynet." Deep learning has been like this so far, because it works in many new domains that have resisted previous techniques like image recognition, speech recognition, go, etc. The "AI Effect" -- saying, "it's just an algorithm" once we succeed -- is an artifact of developing AI by using it as an algorithm for a domain-specific task. It is "just" an algorithm, because that's all it needs to be to win at go, identify cats, or whatever. Fundamentally, it's not very surprising that when you set out to make a system that's really good at playing go...you end up with a system that's really good at playing go. Of course it's hard -- that's what we should appreciate -- but most of the problems we consider "hard" are based on what we think it is hard for a computer to do in order to solve the problem. "Searching the space of all possible go moves is computationally intractable, therefore this problem is hard." In reality, the problem may not be as hard as we think if the agent doesn't have to search the space of all possible moves. "Hard" problems are multi-optimization, where even the definitions aren't very clear: "learn as much as possible and live a happy life while making a productive living for yourself". Turn that into an objective function...

eli_gottlieb 10y ago
>"Hard" problems are multi-optimization, where even the definitions aren't very clear: "learn as much as possible and live a happy life while making a productive living for yourself". Turn that into an objective function... Minimize the informational free-energy of the product of the multiplicatively inverted brain-to-body reinforcement-learned energy function of reward with the brain-to-body reinforcement-learned energy function of punishment, via active inference? Beh, those words are messy. The point is to talk about a Gibbs distribution with energy functions Reward(X, Y) and Punishment(X, Y), such that the "total" energy function is E(X, Y) \propto Punishment(X, Y) - Reward(X, Y). This then gives us a "goal distribution" for an active inference agent (like a human), defined "up to" the reinforcement-learned energy functions, whose limiting functions (need a lot of functional analysis and probabilistic reasoning over function spaces, there) are the "ground truth" causal relations by which the world causes reward signals through the body and its senses. https://arxiv.org/pdf/1503.04187.pdf https://arxiv.org/pdf/1503.04187.pdf

eli_gottlieb 10y ago

>"Hard" problems are multi-optimization, where even the definitions aren't very clear: "learn as much as possible and live a happy life while making a productive living for yourself". Turn that into an objective function... Minimize the informational free-energy of the product of the multiplicatively inverted brain-to-body reinforcement-learned energy function of reward with the brain-to-body reinforcement-learned energy function of punishment, via active inference? Beh, those words are messy. The point is to talk about a Gibbs distribution with energy functions Reward(X, Y) and Punishment(X, Y), such that the "total" energy function is E(X, Y) \propto Punishment(X, Y) - Reward(X, Y). This then gives us a "goal distribution" for an active inference agent (like a human), defined "up to" the reinforcement-learned energy functions, whose limiting functions (need a lot of functional analysis and probabilistic reasoning over function spaces, there) are the "ground truth" causal relations by which the world causes reward signals through the body and its senses. https://arxiv.org/pdf/1503.04187.pdf https://arxiv.org/pdf/1503.04187.pdf