Morphospace / Evidence for intracellular memory storage and its implications for AI

Evidence for intracellular memory storage and its implications for AI

In one paragraph

Glanzman, a neuroscientist who says he has been in the field for over 40 years, argues that memory is not stored at synapses (the connections between neurons) but inside neurons. He says he believed the standard synaptic view for about 32 of those years and changed his mind in the last eight or so. He then draws out what that would mean for AI, where "learning" means changing connection weights, and describes a new system his group proposes in which each unit carries its own memory. In the discussion Levin links it to his own work on memory transfer, anthrobots and the caterpillar-to-butterfly memory puzzle.

Main ideas (plain words)

  • The textbook story is that an experience changes synaptic strength, then gene activity and protein-making build new connections, so memories are held in those changed connections. He cites a survey in which 70.5% of neurobiologists said long-term memories are mainly kept in connectivity patterns and synaptic strengths.
  • He argues that evidence from lab animals suggests otherwise: behaviour and memory can outlast major changes to the connections. He walks through experiments on sea slugs (Aplysia) and on blocking memory "reconsolidation" (the re-stabilising of a memory after it is recalled).
  • He says modern AI inherited the synaptic picture: a trained network stores what it learned in its connection weights, and a standard artificial neuron is stateless.
  • He describes a paper he says is on arXiv (so a preprint, not peer-reviewed as far as the talk says), titled "Neurons are all you need", introducing a system called Naresa ("neural unit with retentive self-adapting architecture"). Each neuron keeps a running memory register and a consolidated "vault" copy, which he says can be moved to an untrained model without retraining.
  • He claims that in the system, neurons choose for themselves whether to take on learning, and that catastrophic forgetting (new learning wiping out old) is reduced. These are his claims about his own work, not independent results.
  • In the discussion Levin says his model is that memories are patterns that can be read and remapped by different bodies, and that what carries over is a high-level goal, not the specific hardware.
  • Levin describes an experiment he wants to try: grow anthrobots from cells taken from smokers and see whether they are drawn to nicotine, which would be a clean case of a memory remapped onto a body with no neurons or muscles. He says plainly "we haven't done this yet".
  • Glanzman proposes that brain activity could release RNA-carrying particles that spread through the body and carry memory-related information. Levin calls it a plausible explanation.
  • Glanzman closes by pointing Levin to a study in which mice trained on fear conditioning, then put into hibernation, lost over half of their hippocampal synapses yet still remembered after waking.

Key quotes (word for word from YouTube's auto-generated captions)

"memory is not stored at synapses but rather is stored intracellularly in neurons"
https://www.youtube.com/watch?v=tzF9C2o1N8c&t=20s

"I would say 32 of those years I believed this"
https://www.youtube.com/watch?v=tzF9C2o1N8c&t=215s

"a standard neuron in an AI system is stateless. It doesn't have memory."
https://www.youtube.com/watch?v=tzF9C2o1N8c&t=3300s

"you can transfer the memory without retraining the model."
https://www.youtube.com/watch?v=tzF9C2o1N8c&t=3405s

"it's not really memory. It's a notepad."
https://www.youtube.com/watch?v=tzF9C2o1N8c&t=3445s

"look anthrobots they don't have any neurons"
https://www.youtube.com/watch?v=tzF9C2o1N8c&t=4450s

"what I tell people is we don't understand anything."
https://www.youtube.com/watch?v=tzF9C2o1N8c&t=4775s

Source: YouTube auto-generated captions, speakers not labeled; wording is provisional. Timestamps point to the start of the stretch that holds the quote, give or take a few seconds. The first three quotes are Glanzman, the nicotine/anthrobot quote is Levin; the "stateless" and "notepad" quotes are Glanzman describing the AI side.