How many patterns can a neuron detect?

[mraptor]

I’ve heard Jeff says that a pattern is detected by 8-15 synapses.
Also that the synapses has to be close to each other and be activated in approximate the same time. Detecting patterns will be physically constrained i.e. reusing synapses for other patterns OR having synapses further away act as a pattern (SDR is not constrained this way)

My question is : On average a 10k connections neuron reacts/detects to how many PATTERNs ?

Second question: if 10% of the synapses are proximal , does the neuron react on the SUM /cummulative/ of the signal on those synapses OR on the PATTERN of it.

A SDR union would probably allow 10-100 patterns.

[jhawkins]

Most of these details are covered in our “Why do neurons have thousands of synapses” paper.
There isn’t a simple answer to your first question. If we said a dendritic spike threshold was 20, then we could simply say 10000/20 = 500 patterns. But there are many variables, so a reasonable answer is “hundreds”.

In our spatial pooler we sum the proximal inputs to each neuron. However, we then pick the top n% winners, and give them a binary activation. Because of the enforced sparsity (k winners) and the binary activation, the neuron acts more like a pattern detector.

[Falco]

Edit: This post entry is wrong. (I leave it in for thread consistency).

@jhawkins

If we said a dendritic spike threshold was 20, then we could simply say 10000/20 = 500 patterns. But there are many variables, so a reasonable answer is “hundreds”.

Wait… that can’t be right.

In ideal situation, 20 connections out of a total of 10000 makes 10000 ! / (20 ! * (10000 - 20) ! ).

That’s too much to calculate. (Previous math was wrong ergo the edit)

If out of the 10k say only 1000 are reliable due to stochasticity and you need a minimum of 8 to have a recognisable pattern, that still would allow 24,115,080,524,699,431,125 different patterns.

And that’s not right either. The synapses need to be close to each other in the so-called dendridic segments. So if a segment has an average of 40 synapses, we need to calculate the amount of segments per typical neuron.

10,000 / 40 → 250 segments

Each segment can produce 76,904,685 unique non-ordered patterns.

76,904,685 times 250 = 19,226,171,250 patterns.

What am I missing?

[jhawkins]

Please read the paper I mentioned. Each small section of a dendrite, approx. 40u in length acts like an independent coincidence detector.

[mraptor]

the question is not how many patterns you can generate, but how many you can detect (given size and sparsity).

In computer terms imagine a SDR-union the capacity is ~10-100 patterns detection … but the same size/spa specs you can express gazillion patterns.

The same goes for the neuron it have a detection capacity constrained by physical properties

[Falco]

Ok, I think I get it now. My apologies, I was confused.

In the neuron paper, (which I had read twice, and still managed to misunderstand) on page 3 right column is written:

The synapses recognizing a given pattern have to be co-located on a dendritic segment. If they lie within 40μm of each other then as few as eight synapses are sufficient to create an NMDA spike (Major et al., 2008). If the synapses are spread out along the dendritic segment, then up to 20 synapses are needed (Major et al., 2013).

and

If we assume an average of 20 synapses are allocated to recognize each pattern, and that a neuron has 6000 synapses, then a cell would have the ability to recognize approximately 300 different patterns.

The thinking error I made is that with 6000 synapses (in this example) billions of different combinations can be tested, but only for 300 out of those the neuron will actually fire.

Sorry again.

[mraptor]

nothing to be sorry about it is complicated thing with many moving parts