The Synapse Memory Doctrine Threatened? | Hacker News Reader

The Synapse Memory Doctrine Threatened?

blogs.discovermagazine.com

51 pointsDiabloD38 comments

Followup experiments with more than two neurons should be interesting, to see whether graph arc destinations are remembered, or just the total.

A far-fetched mechanism could be that each neuron transmits a signature (UID), that we've not detected because we've not been looking for it (perhaps it's slower than expected, taking hours to transmit it once...). It could include gross addressing information, based on chemical gradients at the time of its formation.

I'm partial to the idea that neurons use more than just synaptic connections. It just seems too simple, especially when there's all that powerful cellular machinary (e.g. RNA transcription) just lying around...

This is certainly interesting but it doesn't look conclusive

How will this impact research into artificial neural nets?

Artificial neural networks are very distant from actual neuroscience. Researchers try to make their nets more "brain-like" but it rarely results in better algorithms. In the end they just use whatever works, and that happens to involve a lot of algorithms which are "biologically implausible". I.e. things that the brain couldn't physically do even if it wanted to.

E.g. sharing learned weights between different parts of the net is extremely effective, but probably not possible in actual physical neurons. Even the main algorithm used to train NNs, backpropagation, is considered biologically implausible. Although a lot of theories exist as to how neurons could implement very similar algorithms.

I think that gasnet neural networks that took into account the locally inhibitive? impact of nitrous oxide are far superior to traditional ANN. But I just read that recently, so maybe misunderstood that.

To put it another way, while it’s easy to see how a neuron could ‘store’ a scalar variable using epigenetics, it’s much harder to imagine that it could store a vector of values.

Of course a vector is just a scalar in disguise. Take Gödel numbering and you can store anything, even if all you have is unary coding.

Given the complexity of the genetic code, why would we just consider this implausible from the outset?

Depends on your system of computation. If all you can do to your numbers is add, multiply, but you don't have a eg a floor function, you can't do Goedel numbering.

It's way too complex, doesn't take into account noise, etc.

Reply on news.ycombinator.com