Hebbian learning is shown to work through the number, not size of synapses

I guess this confirms HTM’s binary synapses?

8 Likes

Subutai pointed out this paper on twitter.
And now a word from the analog values camp …

Abstract

The active electrical properties of dendrites shape neuronal input and output and are fundamental to brain function. However, our knowledge of active dendrites has been almost entirely acquired from studies of rodents. In this work, we investigated the dendrites of layer 2 and 3 (L2/3) pyramidal neurons of the human cerebral cortex ex vivo. In these neurons, we discovered a class of calcium-mediated dendritic action potentials (dCaAPs) whose waveform and effects on neuronal output have not been previously described. In contrast to typical all-or-none action potentials, dCaAPs were graded; their amplitudes were maximal for threshold-level stimuli but dampened for stronger stimuli. These dCaAPs enabled the dendrites of individual human neocortical pyramidal neurons to classify linearly nonseparable inputs—a computation conventionally thought to require multilayered networks.

3 Likes

Those action potentials are per dendrite, not per synapse. I remember them described as analog version of XOR operator? Maybe some kind of backflow in dendrite segment?

I’m always puzzled by the absolute need of a lot of peoble to explain the brain in terms of logic gates.
What is so hard about Hebbian learning?
Thats all a synapse can do.
XOR is: x y || z

         1 1 || 0
         0 1 || 1
         1 0 || 1
         0 0 || 0

All agree that a single synapse on a dentrite does nothing.
So you would have lets say at least 10 synapses on a dentritic segment.
XOR on that means that any permutation of 5 zeros and 5 ones will give a 1, the rest a zero.
Who actually believes that synapses work that precise???
We know they don’t! So it does not help very much to think of it in terms of logic gates.
Because 4 ones and 6 zeros is not almost 1, its zero! But i’m pretty sure in the brain its sometimes one.
Also the distal dentritic segments are not an AND gate!
It’s a treshold.
Because with an AND all the synapses on a segment need to be active,
with a treshold one or two can be missing.
And so on …
We know that those Calcium spikes are a bit stronger and longer than NMDA spikes.
So how about that: they do the same biasing as the NMDA spike to enable the temporal memory,
just that this signal comes from above
so a higher region has “understood” something and now it “tries” to put the lower hypercolumn in agreement if at all possible
enabled by the fact that they can potenialy overwrite an NMDA spike
Just a hypothesis, but certainly more plausible than an XOR.

It’s doesn’t work like XOR, I did say an analog version. But it can perform XOR if you reduce output resolution to binary. NMDA is a receptor, I think it’s also a main calcium pump, what other calcium pumps are there?
Your proposed mechanism would only work in specific cases, depending on interaction between regions.
The grading described in the paper is general, so it should be a product of local interactions, regardless of where the inputs come from?

To be clear, the depolarisation or biasing of a neuron to make it fire faster in the HTM is modeled after NMDA spikes on distal dentrite segments,

subutais paper is not very clear on that, but i read it as talking about apical dentrites, and they do get activated from higher regions

Ok, I meant the paper posted by Bitking: https://science.sciencemag.org/content/367/6473/83

Yeah, i meant the same paper, subutai posted it on twitter :wink:

I read that paper as talking about apical dentrites

So, even if the inputs are from higher regions, the grading is in output of dendrite segment, not the whole neuron. Which means it’s produced by interaction between synapses / potentials within a segment, not between apical and proximal inputs?.

Thats true, its a property of those human L2/3 apical segments

On a very high level i think of it like this:
the temporal pooler is a sequence memory that accesses itself with those distal dentritic NMDA spike
so dentritic spikes are like a address operation to preselect a pattern
how about those apical spikes do the same thing, just from above
irrespective of the technical details
apparently one needs to limit the activation from above, ok, thats fine
it’s still unlikely to just be an XOR, because single neurons are never that important or precise
the ensemble of apical synapses could do “memory access” to pre select a SDR in L2/3
just like in the temporal memory

Maybe clock signal is a better metaphor than address operation. Clock may have to be exclusive, to avoid ambiguity in the output. But I would think axonal hillock is a far more logical target for a clock than apical dendrites. Just my speculations.