4 ms·
> have detailed how memory is encoded by synaptic weights Synaptic weights is our computational model of neural networks that seems to work pretty well. But th
by titzer 6y ago
> have detailed how memory is encoded by synaptic weights
Synaptic weights is our computational model of neural networks that seems to work pretty well. But this is a model. No one has discovered the biological mechanism by which synapses are "weighted". The actual computation going on in a cell is totally based on chemistry.
There are a lot of different specialized networks in human and animal brains that don't necessarily translate easily to "weighted synaptic networks" all that well. For example, networks that process audio information have been shown to recognize individual frequencies (in fact, they pretty much solely do this!) Such networks seem to be primarily based on the firing rate across a synapse. For such networks is seems more reasonable to think of neurons as signal processors whose internal computation is more based on the time domain, e.g. some kind of Fourier transform-equivalent.
In short, I think you should put less weight (pun intended) in our computational model of neurons (link weights) and look more at biology.
- jrussino 6y agoIt has been over a decade since I studied neurobiology, but from what I recall we do have at least some understanding of how synapses become "weighted". For example: long-term-potentiation, the idea that "neurons that fire together wire together" https://en.wikipedia.org/wiki/Long-term_potentiation https://en.wikipedia.org/wiki/Long-term_potentiation https://www.khanacademy.org/test-prep/mcat/behavior/learning-slug/v/long-term-potentiation-and-synaptic-plasticity https://www.khanacademy.org/test-prep/mcat/behavior/learning...
- netizen-9748 6y agoLTP is only one mechanism. Neurons vary their responses in multiple ways, usually some form of receptor protein modification. But again, this is one mechanism of many.
- wombatmobile 6y agoPerhaps mRNA are performing the synaptic weighting. After gene expression, mRNA's are left floating as debris in the cytoplasm, until they are cleaned up and expelled. If the gene expression continues chronically, the same mRNA sequences will populate the same cells, transiently, for as long as the gene expression recurs. What if the mass of the mRNA sequences influences the electrical conductivity of the cell e.g. functions as a form of antenna. In the aggregate, cells with mRNA populations e.g. neutrons active in particular thought patterns, might have stronger antenna, or differentiated electrical characteristics, and hence influence thought more or less in the network, for the duration of the gene expression.
- subroutine 6y agoSure, debris mRNA may affect neuron electrical potentials. But this would be an extremely wonky way for neurons to regulate electrochemical propagation. One that has not been observed irl, afaik. However we know a ton about how neurons directly regulate electrical potentiation by modifying the expression of postsynaptic neurotransmitter receptors. The mechanism is simple... the more receptors in a given synapse, the greater the dendritic membrane will depolarize from each upstream signal (signal being each synaptic release of a neurotransmitter vesicle by the upstream axon). Electrophysiology instruments are so sensitive now-days that we can detect, in a quantized way, the release of single synaptic vesicles. We can then do things to increase synaptic expression of neurotransmitter receptors, and find that each vesicle release results in a greater depolarization of the downstream neuron. Long story short, the quantity of neurotransmitter receptors at postsynaptic membranes is a physical correlate (probably the primary physical correlate) of synaptic weights in artificial neural networks.
- pontifier 6y agoYour comment sparked a thought about a possible collection of RNA memory functions: If RNA is the memory mechanism, and it floats around for a while, then it gets cleaned up over time, this collection might allow long term memory recall and replay. Function 1: Surface protein encodes the pattern of incoming synapse firings on RNA, then releases it. This pattern would have a start and end block indicating it is "new". Function 2: Relatively rare protein finds RNA marked "new" and replaces start and end blocks with "old" markers. Function 3: Common protein grabs one strand marked new, and one strand marked old and compares part of the tail of "new" to the head of "old" by sliding them past each other. If a match is found, the older strand is marked "match" at the start. Function 4: Match strands are used to trigger the next several firings of the neuron, then marked "hold" Function 5: "Old" strands are gradually recycled. Function 6: "Hold" strands are copied into new "old" strands for comparison to "new" strands by function 3 Function 7: Something to do with continual refinement of "hold" strands with incoming "new" strands. ... Anyway, this whole little system just popped into my head based on the small amount I know about crispr and cellular machinery (which honestly isn't much), and I wanted to share.
- subroutine 6y agoWe have discovered it, and understand it quite well. It's fairly simple, at the level of the synapse there are ultimately three ways to evoke increased firing in the downstream neuron: (1) the upstream neuron's axon can release more neurotransmitter, (2) the upstream neuron can establish more axon collateral connections with the downstream neuron, or (3) the downstream postsynaptic membrane can express a greater number of neurotransmitter receptors.