3 ms·

>"And it’s where we should all be looking: gradient descent plus symbols" how about no gradient descent at all

• lkrubner 8y ago
  The problem is the exponential explosion of possibilities. Fairly amateur Python code can work through a problem in 10 minutes which, without gradient descent, would literally take a thousand years.
  • YeGoblynQueenne 8y ago
    The term is "combinatorial explosion". Like the ability to train larger, deeper, heavier neural nets, that too gets better with more computing power.
  • bra-ket 8y ago
    our brain is not doing brute force search
    • eli_gottlieb 8y ago
      Right, it's doing stochastic search through a Bayesian posterior distribution. The trouble is that established inference techniques can't really handle very general, "human-level" model classes in real time, computationally.
      • throwaway2048 8y ago
        We have no idea if that is true or not.