> 1. Most neural network architectures have fewer "neurons" than 10^5. Maybe the word you are looking for is "parameters"
That's a fair point, I shouldn't have said "typically." But some of the larger models probably have that many linear filters.
> 2. A neuron in the brain and a neuron in a neural network are totally different things.
There are certainly disanalogies between biological neurons and neurons in a vanilla feed-forward network, but A) there are analogies as well and B) a lot of interesting work is being done to make deep learning models stdp compatible.
At any rate, I think it's a reasonable claim that an insect brain has representational power closer to a SOTA ANN than it does to a human brain (though I welcome anyone here who knows about biologically plausible deep learning and/or insect brains to prove me wrong)
I entirely disagree. The onus of proof is on you to explain how a highly idealized ANN has the same expressivity as a fly's wetware.
Just because neural networks are tech's Zeitgeist, doesn't make them the perfect explanation for all physical phenomena. Fifty years ago there were people equating thought and consciousness with artificial intelligence programs; and 200 years before it was the watchmaker's clockwork holding that regard.
Yes, as a man of science I subscribe to reductionism. Brains are made of neurons which are made of molecules which are made of atoms and so on, all governed by the laws of physics. But there's no reason to believe ANNs have the required intrinsic complexity to behave like ganglia.
> The onus of proof is on you to explain how a highly idealized ANN has the same expressivity as a fly's wetware.
I've made an argument - it's essentially a functionalist one. The intelligent things insects do - object detection, maze navigation - are all things that ANNs are really good at.
To put things in perspective: the reason that I don't think that ANNs are anywhere near as expressive as the human brain is that there are countless behaviours that humans perform that ANNs simply can't - generalizing to novel viewpoints in vision, for example. (Or if you want to go whole hog, natural language understanding.)
AFAICT The same is not true of insects. To dissuade me, you'd have to specify an insect behaviour that ANNs are fundamentally unequipped to perform. I'm not an entomologist so I'm totally open to the possibility that there is one. I also imagine there is significant neural diversity within the insect kingdom - presumably some bugs are smarter than others, and maybe some of the bigger-brained ones can do stuff that would necessitate an explanation that invokes consciousness. But you have to tell me what it is.
That's a fair point, I shouldn't have said "typically." But some of the larger models probably have that many linear filters.
> 2. A neuron in the brain and a neuron in a neural network are totally different things.
There are certainly disanalogies between biological neurons and neurons in a vanilla feed-forward network, but A) there are analogies as well and B) a lot of interesting work is being done to make deep learning models stdp compatible.
At any rate, I think it's a reasonable claim that an insect brain has representational power closer to a SOTA ANN than it does to a human brain (though I welcome anyone here who knows about biologically plausible deep learning and/or insect brains to prove me wrong)