EXACTLY! So many hung up on exact neural correlates. That work needs to continue, and it will, but things that work are always welcomed by engineers.
I like to think of it like the speculations on life itself; i.e., other than carbon-based life forms.
It’s actually quite easy. I’m surprised you’re sceptical about it. I’m quite far beyond that point now. I’ll be publishing the work on github soon when the work is complete, along with comprehensive documentation.
Unlike biology, we have tricks up our *sleeves. We can optimise algorithms and hardware for a specific function. Biology is far too slow to evolve these solutions. The cortex may have trillions of synapses, but only a small percentage of them are used during inference and our algorithms can cut the cost of summation and pooling by using hashmaps as a very basic example - a luxury biology doesn’t have.
Biology has its bag of (still) magic tricks too. Otherwise we would have grounds to claim we understand it.
It is possible a very small proportion of neurons (mini columns maybe) are those which “know” anything/everything and the massive majority are needed to handle the unknown, the not understood and the couldn’t predict it.
Like a kind of search engine but not for records within vast archives but searching for needles of sense/meaning in the haystack of (relatively) recent experience, which gets summoned whenever the minority of “expert” columns fail to account for that experience.
As Piaget put it
Intelligence is what you do when you don;t know what to do
.
I’m not a computational biologist, so my opinion means nothing. BUT… I do have an opinion - it’s cheaper not to propagate at all - but to ‘latch’/associate onto activate states/assemblies. This mechanism is already there and ready to go. It’s the structure of the associative & gating machanism that eliminates the need to do backward/forward propagation at all - it’s real time learning.
Sorry it’s been a while. It’s getting there. It’s like building an assembly language for an associative memory machine. The space of alternative computation is vast and exciting… but frustrating because it takes a long time to build a house.
Dendritic backprop is inside a neuron: [2211.11378] Learning on tree architectures outperforms a convolutional feedforward network, so it uses all these active states.
Cheers Boris for the link. I wouldn’t be suprised that biological mechanism will needed to propagate a signal from one endpoint to another, even with state is needed to determine the endpoint. I was teasing at the idea that assemblies of neurons in different units assemble together in quorums to encode the state. This being just a small step in the whole cycle.
I think you mean Neuronal ensemble - Wikipedia. AFAIKT, they are mediated primarily by laterally forking axons, in layer II-III. There is a short-range lateral inhibition and long-range excitation (“Mexican hat”), and the later cause reverberation and “pattern completion”.