As a former ecology student and EA city group organizer, this is the perfect confluence of topics for me. It reignited a slumbering interest in these topics for me, and discovering that the sperm poison thing wasn't just a toy example but a real thing later in the article was *so* fascinating. Is the article supposed to suddenly cut off at the end, or do I need to become a paid subscriber or something?
I really think there probably isn't an analogy to recombination currently in ML. Recombination depends on there being homologous copies of each chromosome in the organism and in the larger population. There's nothing like that with the neural net parameters.
The closest thing would be if there are circuits that are analogous the genes and there were some reassortment mechanism to keep them from forming cabals that don't serve minimizing overall loss (like the poison-antidote pairs). I suppose such a thing could evolve within a model by gradient descent, but if so we don't know about it.
As a former ecology student and EA city group organizer, this is the perfect confluence of topics for me. It reignited a slumbering interest in these topics for me, and discovering that the sperm poison thing wasn't just a toy example but a real thing later in the article was *so* fascinating. Is the article supposed to suddenly cut off at the end, or do I need to become a paid subscriber or something?
I fixed it! Thank you so much for letting me know, and please enjoy the end :)
Oh shoot it kept giving me a warning about the length, but I thought I had fixed it. Looks like it's truncated...
Very interesting. What would the AI parallel to recombination be?
I really think there probably isn't an analogy to recombination currently in ML. Recombination depends on there being homologous copies of each chromosome in the organism and in the larger population. There's nothing like that with the neural net parameters.
The closest thing would be if there are circuits that are analogous the genes and there were some reassortment mechanism to keep them from forming cabals that don't serve minimizing overall loss (like the poison-antidote pairs). I suppose such a thing could evolve within a model by gradient descent, but if so we don't know about it.