3 ms·

Nueral network tech went through a weird period and then back to basics (back prop) just with massive computation (and simple activation functions) with fabulou

• veets 6y ago
  There is plenty of work running evolutionary algorithms on GPUs, but there are a number of limiting factors that make it less than ideal. Evolutionary algorithms in general can be extremely effective, they just solve very different problems from "is this a picture of a cat?" and aren't as sexy as neural networks. In fact, much of "old school" CS is still very much alive and applying modern techniques and hardware to their problems. It's not that neural networks are some field of CS that is much more advanced than the rest of CS but rather neural networks happen to be the current hot topic.
• imvetri 6y ago
  I think GA with artificial neuron will be a good combination.
• djaque 6y ago
  What do you mean a massive GA? Just using a huge population? The issue is that in a GA, time is dominated by evaluation of the fitness function. For my applications every time you do that takes ten minutes. More individuals just means more cost. Plus now there are more of them randomly searching or doing the same thing as each other. So in a sense there is a "too big".
• matsemann 6y ago
  For GAs the expensive part is often the fitness function. So it's very problem dependant.