Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Evolution Is the New Deep Learning

sentient.ai

Post reply on HN
Print view View original

41–50 of 242 posts

Re: Evolution Is the New Deep Learning

#41

Post by klmr »

blacksmith_tb wrote: post #12
XnoiVeX wrote: post #6

A philosophical tangent: Evolution did not always produce the best outcomes. Would neuroevolution be vulnerable to similar effects? Probably a good research area. https://www.wired.com/2009/07/st-best-5/

A can think of a few objections to that argument. First, evolution isn't 'over' for any of the species on that list, and if any of those seemingly maladaptive traits really do cause serious problems for those species, we would expect them to be selected against (there are some cases where they are associated with other traits that are important and helpful, and so won't be tossed overboard even then). Evolution also…

I agree, the examples aren’t the best chosen. But there are cases which are pretty well established in biology as being maladaptations. — The laryngeal nerve is the most famous of those [1] because there’s really no advantage to the way it develops, and its characteristics poses significant health problems that do affect evolutionary fitness negatively. Yet there’s no easy fix to solve its positioning through gradual mutations.

[1] https://en.wikipedia.org/wiki/Recurrent_laryngeal_nerve#Evid...

Re: Evolution Is the New Deep Learning

#42

Post by maksimum »

xants wrote: post #31
yorwba wrote: post #29

Earlier quoted context omitted.

If you mutate genomes by small additive modifications to a vector of continuous parameters, then taking lots of samples and keeping the best is essentially a stochastic approximation to gradient descent. However, unlike the SGD used in deep learning, it doesn't make use of calculus and therefore requires many more samples (exponentially more, in the worst case) to get a gradient of equivalent accuracy. I.e. it's slow…

Is it true that genetic algorithms have the benefit of being able to find the global optima more consistently due to the incorporation of randomness in subsequent generations? Whereas DL models often get stuck at local optima?

What do you mean by "find" and how do you define "global optima"?

For "find" you could discuss convergence rates vs. the points that are being converged to. If you add randomness at every iteration you're not even converging, at which point annealing rate becomes a related issue.

For "global optima" are we talking about training error, or test error, or some other kind of function value?

Re: Evolution Is the New Deep Learning

#43

Post by maksimum »

TheOtherHobbes wrote: post #34
boffinism wrote: post #8

Interesting to see that we're a long way from this technique actually generating anything valuable. E.g. if you try the LSTM Music Maker they link to ( https://www.sentient.ai/sentient-labs/ea/lstm-music/ ) and enter a melody, the resulting 'improvisation' will neither have any of the hallmarks of your initial input nor obey any of the conventions of any genre of music I recognise. It'll just spew out a random-seemin…

Evolutionary approaches won't necessarily produce better music. This is a well-researched area. The challenge with all GA work is finding the right fitness function. If you don't understand your domain well enough to define a good fitness function you're going to be wasting your time, and no amount of algorithmic and/or AI magic is going to help you.

I think the point here is that we don't need to define a fitness function for "good music", but a fitness function for e.g. "reconstruction error" for auto-encoders. While the latter is still challenging, it seems slightly easier, although I'm not sure that it's significantly easier than general manifold learning...

Re: Evolution Is the New Deep Learning

#44

Post by jorgemf »

mav3r1ck wrote: post #15

Having studied this extensively back when they were called Genetic Algorithms, I would like to offer a few insights. 1) One of the biggest reasons they fell out of favor for more "mathematical" approaches was that no one could really explain why exactly they worked. It makes sense on the surface that "survival of the fittest" and doing something akin to multiple stochastic gradient descents would work, but no one has…

Regarding your points:

1) "explainable AI" is better in GA than in deep learning. GA gives you a structure that works and probably easier to understand than any Deep Learning model (which it is a huge math function). There are so many things that also doesn't make sense why they work in deep learning but we still use them, that is the same with GA.

2) Knowing the fitness function doesn't mean you can solve the problem. When the search space is so big you need something to search on it, and there is where GA can shine. It is also the same with Deep Learning and mostly evolutionary computation. The search space is so huge for a "brute force algorithm". You need heuristics and GA works well for some problems, the same way gradient descent works well for others too.

Re: Evolution Is the New Deep Learning

#45

Post by sabertoothed »

cvaidya1986 wrote: post #11
XnoiVeX wrote: post #6

A philosophical tangent: Evolution did not always produce the best outcomes. Would neuroevolution be vulnerable to similar effects? Probably a good research area. https://www.wired.com/2009/07/st-best-5/

What’s the definition of best

Not having a 5 meter detour [1] of a nerve.

[1] https://en.wikipedia.org/wiki/Recurrent_laryngeal_nerve

Re: Evolution Is the New Deep Learning

#46

Post by verdverm »

chrisfosterelli wrote: post #21

These "evolved neurons" look ridiculously more complicated and expensive to evaluate than a base neuron for a minor improvement in performance. The Music Maker paper mentioned that LSTMs haven't changed in "25 years", but didn't mention the 2014 Gated Recurrent Units that have become so popular specifically because they are much more simple to understand and evaluate. This approach seems to be going in the opposite d…

Likely solution bloat, another of the EC/GP oddities that arise. During reproduction you randomly disrupt solutions, so to protect good solutions from being disrupted, the solutions increasingly add irrelevant stuff. This drops the probability that the good part gets messed up, but leads to loss of population diversity, as there are an infinite num er of useless things you can add /d o

Re: Evolution Is the New Deep Learning

#47

Post by simion314 »

IAmEveryone wrote: post #40

If only wishing made it so... Evolutionary approaches have always had one big feature in their favour: they are far more fun to work with. They produce all these fascinating oddities, like the one that learnt to outsmart it’s opponents at infinite tic-tac-toe by playing coordinates 10 2312 and 4 7875 and watching them run out of memory trying to build a data structure for the board. They are also far easier to combin…

As others mentioned the hard part is modeling your problem, how to encode the "individuals", defining a good fitness function etc. I think he neural networks are more poplar now because are easy to use, you throw data at it and you get some results(though IMO I am scared that we are using things that we do not understand exactly how they work)

Re: Evolution Is the New Deep Learning

#48

Post by sampo »

mav3r1ck wrote: post #15

Having studied this extensively back when they were called Genetic Algorithms, I would like to offer a few insights. 1) One of the biggest reasons they fell out of favor for more "mathematical" approaches was that no one could really explain why exactly they worked. It makes sense on the surface that "survival of the fittest" and doing something akin to multiple stochastic gradient descents would work, but no one has…

> no one could really explain why exactly they worked

Genetic algorithms work on problems where some subsets of variables are approximately separable (uncorrelated) from some other subsets of variables.

F(a,b,c,d,e,f) ≈ F1(a,b,c) × F2(d,e,f)

So if you have found a good combination or a,b,c it makes sense to try how it works it any promising combinations of d,e,f. Some natural world problems really have this property.

Re: Evolution Is the New Deep Learning

#49

Post by CretinDesAlpes »

billconan wrote: post #35

the music generation demo isn't really impressive to me. the outputs were pretty random.

Yes, this is really bad, not sure what they want to show.

"The AI is processing your melody...", sure, but it can be processed without AI and you could get a much better output. At least please keep the output notes on the same input scale...

Re: Evolution Is the New Deep Learning

#50

Post by verdverm »

mav3r1ck wrote: post #15

Having studied this extensively back when they were called Genetic Algorithms, I would like to offer a few insights. 1) One of the biggest reasons they fell out of favor for more "mathematical" approaches was that no one could really explain why exactly they worked. It makes sense on the surface that "survival of the fittest" and doing something akin to multiple stochastic gradient descents would work, but no one has…

If you like GP, you might like PGE more (self bias)

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.394...

Post reply on HN