What Causes New Neurons to Form?

Hello All,

I’m an artificial intelligence developer. I’m working on a new type of AI. Neural Networks are getting increasingly contrived and specific. My goal is to return to first principles of biology and evolution, creating “more human” Neural Networks.

I’ve built an eye and Occipital Lobe (its still early) that is more similar to the human visual processing system than anything else I’m aware of. It has Rods, Cones, Bipolar Cells, Ganglion Cells, and more… I’m now working on incorporating Horizontal and Amacrine Cells.

As for the “processing” I’m currently focusing on the principles of the Occipital Lobe - ie. “Am I looking at something of interest?”. I’m “growing” the Occipital Lobe from stimulus. Meaning, the Occ. Lobe is generated in training (generation/degeneration of neurons and synapses).

Whenever I’m presented with a new problem/need I turn to Bio/Neuro for a solution, rather than creating a contrived, single-use solution. This means I mainly focus on creating an environment of first principles, that guide the generation of the Neural Network. ie. “Stress, Recovery, Adaptation” being a first principle.

What I’m struggling with is - What stress/signal causes new neurons to form? I’m sure this is a complicated answer… but again, remember, I’m looking for first principles.

ie. Right now I have my Neural Network following a basic “resource limitation” generation structure. (1) Grow Synapse size first, (2) Generate new Synapses second, (3) Generate New Neurons last. Neurons are created when “more signal” is needed, and new Synapses are already very large and plentiful.

Meaning, the stimulus for a new neuron is “signal bottlenecking”.

“What stress/signal causes new neurons to form?”

Thank you, All!

Here’s a review paper (https://doi.org/10.1016/j.neuroimage.2018.12.043) that covers the various developmental stages. From a quick skim, I don’t see much indication that neurogenesis plays a significant role in experience-dependent development of occipital cortex. Plasticity and pruning are likely to do more of the work.

There is quite a bit of existing theoretical/computational work in modeling visual cortex, a lot of which came in the 80s from Marr, Grossberg, Hopfield, von der Malsburg. Some things fall out of geometry and local excitatory/inhibitory connections (such as on-center-off-surround), and adding in Hebbian learning laws can allow for plasticity. IIRC, self-organizing maps were developed out of similar explorations.

Hey @effigies! Thank you so much for this reply. This review paper seems great. I’m working on getting access to it.

My current work is to revitalized the fairly dated work in self-organizing maps. Thanks for your pointers here!