Why Neurons Have Thousands Of Synapses, A Theory Of Sequence Memory In Neocortex


#1

By @jhawkins and @subutai

Available on Frontiers in Neural Circuits and arxiv. Please discuss this paper below.


Need references on cortical mini-columns
Neuroscience of Layers 2/3
HTM Learning algorithm
#2

Hi @jhawkins and @subutai,
I was thinking about hierarchy of regions and how they communicate. One simple problem is higher region has slower pattern that lower region receives and create more complex and faster sequence. So my thinking is
1. apical dendrites receives pattern from higher region and put cells in predictive state
2. distal dendrites receives pattern from same layer that was previously learned
3. usually distal dendrite will put cell in predictive state but if it was already there by apical dendrites it will be activated
4. activated cells sends information to other cells by distal dendrites and if they were already in predictive state by apical dendrite cells goes to active.
In that way sequence unfold by context that is controlled by higher region. Please correct me if I’m wrong. Maybe there is another way that sequence can unfold in some context from higher region.


#3

Hi
That’s pretty close to my understanding but I believe it is a bit more complicated than that.

First, the biological evidence as far as I know is that pyramidal neurons in different layers behave differently. In layer 3 the evidence suggests that simultaneous activation of apical and distal dendrites is not capable of causing the cell to spike. In layer 5a the evidence suggests the it is possible to generate a spike by activating both apical and distal dendrites.

I have a theory for why this is. Activity in layers 4, 3, and 2 always reflects sensory input. If these cells are firing that means something is really happening in the world. If layer 3 learns sequences we don’t want it to be active if no input is occurring. That would be like an hallucination. Predictions in layers 4 and 3 are always represented by depolarization and therefore you are not consciously aware of these predictions. Only if the input doesn’t match the prediction will you be aware of the discrepancy.

Layer 5a is the motor output layer of the cortex. When we generate motor sequences, layer 5a cells are the ones doing it. Learned motor sequences need to go forward on their own. As a motor sequence plays out the cortex verifies that the expected result occurs but the motor sequence itself cannot be driven by sensory input. E.g. imagine speaking a word. Your layer 5a cells drive the voice box but they need to do this before hearing the results of you speaking. This is why layer 5a cells need to be able to play back sequences on their own, without proximal input. L5a cells need to learn motor sequences and they do using their proximal synapses, but unlike layer 3, then can play back sequences without proximal input.


#4

Might there not be variation on this. The vividness of mental visualization varies in the human population, or so I’ve heard. I’ve heard some claim they can vividly visualize things to such a a degree such that they can’t distinguish recall of their memories or imagination from real world experiences.


#5

I am curious if you see this as an implication that for motor commands to be executed, a hierarchy of at least two regions is required (i.e. a region by itself cannot drive motor commands)? Or is there likely to be another process by which a single region by itself can turn predictions into actions? Sorry if that is a dumb question – I am definitely not at the level of comprehension on this subject as others on here (this particular point is relevant to an experimental app I am playing around with)


#6

That’s an interesting question. I hadn’t thought about that before. I would prefer that two regions arranged hierarchically would not be required because as a general rule I try not to rely on hierarchy to solve a problem like this. But I don’t know. I suspect the answer will come with a more thorough understanding of exactly how a region learns behaviors related to objects. [quote=“OS_C, post:4, topic:572, full:true”]
Might there not be variation on this. The vividness of mental visualization varies in the human population, or so I’ve heard. I’ve heard some claim they can vividly visualize things to such a a degree such that they can’t distinguish recall of their memories or imagination from real world experiences.
[/quote]

There might be variation, sure. While it is well known that our memories can fool us, not being able to distinguish imagination from real world experiences is a pathological condition (aka hallucination).


#7

There might be variation, sure. While it is well known that our memories can fool us, not being able to distinguish imagination from real world experiences is a pathological condition (aka hallucination).

I think it can be considered hallucination if the individual experiences them with their eyes opened and interacting in either a controlled or uncontrolled manner with incoming sensory input. IF they close their eyes, and can generate from memory past events vividly, and they know they’re actually remembering, I think such wouldn’t necessarily qualify as abnormal even if it was very hard to distinguish or indistinguishable in a qualitative sense from external input.

But I’ve not researched whether those with high scores on say the “vividness of visual imagery questionnaire”, whether these individuals are more prone to hallucination and mental illness.


#8

Thanks @jhawkins,
Your answer clears why sequence unrolling is not possible in layer 3 that is processing input information. So i guess that I can conceptualy separate upward and downward layers

  • layers 4, 3, 2 are used for upward processing
  • layers 5, 6 are used for downward processing
  • feedback between upward and downward processing
    Please correct me if I’m wrong.

#9

That’s basically correct. But we should be careful because the connections are more complicated than that. The upper layers do get feedback. Layers 3 and 2 receive feedback from higher regions. This can bias those layers to interpret the input in a certain way. But I don’t think the feedback can activate the cells on its own. E.g. if I say “do you see the dog in the cloud” you are much more likely to find a dog image in the cloud because you are biased to look for a dog (I believe this is due to feedback to L2), but you won’t see anything if you aren’t looking at the cloud.

Layers 5 and 6 pass information down the cortical hierarchy, that’s a fact. If I ask you to sign your name activation has to flow down the hierarchy towards M1 and M2. This goes from L5 to L6 to L5 (via apicals in L1) to L6 etc. The feedback must activate cells in L5 and L6. However, part of L6 receives receives FF input and it passes that to L5. What is going on here is what I am currently working on and writing for a journal paper. It has to do with coordinate transformations and we are still working through the details.


#10

Thanks @jhawkins for your anwser. You explained a lot to me. There are two things that I’m thinking:

  1. What is sent to upper region
  2. How we learn by sending information down the regions (hipocampus -> higher region -> lower region)
    For first I was thinking that cells from 2/3 layers sends information to upper region. If pattern is not predicted there will be greater density of active cells what will lead to more information passed to upper region. For second I haven’t find anything that makes sense to me. Maybe you could add something to it. Please correct me if I’m wrong for first one.

#11

@jhawkins I disagree. Top-down influence via L1 apical dendrite inputs must give rise to cellular activation in L2/3 at least some of the time, there is no other way to explain many behavioural phenomena.


#12

Hi Fergal,
What behavioral phenomena are you thinking about? If I am wrong about this I want to know.
Jeff


#13

@jhawkins I was thinking what is passed to upper region

  1. Predicted cell that is activated in 2/3 layer inhibits other cells nearby by fast inhibition
    a) There are less active cells in predicted patterns (low density of impulses)
    b) Active cells (low density) from 2/3 layer sends information to 4 layer in upper region
    c) There will be low density of information in 4 layer in upper region for predicted patterns
    d) Low density will lower probability of activation of 4 layer cells in upper region
    e) Less information will be passed to 2/3 layer of upper region
  2. Unpredicted pattern will not have cells in predicted state (no inhibition)
    a) There are more active cells in unpredicted patterns (high density of impulses)
    b) Active cells (high density) from 2/3 layer sends information to 4 layer in upper region
    c) There will be high density of information in 4 layer in upper region for unpredicted patterns
    d) High density will increase probability of activation of 4 layer cells in upper region
    e More information will be passed to 2/3 layer of upper region.
    It is simple concept that doesn’t add anything new to the table. That are just known concepts that are combined together. Does it makes sense for you?

#14

Hi Marko,
I am a bit confused. My question was what is the evidence that top down feedback projections to L1, from an upper region, makes cells in L2 or L3 active? I expect top down feedback to L1 will depolarize cells in L2 and L3 but not make them active. I also expect top down feedback to L1 can cause L5a cells to become active. But I don’t expect top down feedback to L1 will cause L2 or L3 cells to become active. I think this is what Feral was arguing for and I wanted to know why he thought that.

You seem to be discussing feed forward projections only. Did I miss something.

BTW, my current thinking about FF projections differs from what you wrote. 1) I assume that FF input to a region goes through a spatial pooling process. The SP always outputs a set of active columns with the same sparsity. For the most part the sparsity of the input doesn’t effect the sparsity of the region. 2) There are two FF inputs to a region. One is a direct L23 input, the second is a L5a to thalamus to L4 pathway. Sherman Guillery argues that the second path through the Thalamus is the main feed forward pathway.
Jeff


#15

Hi @jhawkins,
Thanks for your answer. On the latest post was thinking about FF projections. I guess that spatial pooling is done by layer 4 cells in cortex. I was thinking to explain why high order changes (unpredicted patterns) are passed to upper layer. High density of active cells in unpredicted pattern in 2/3 layer made sense to me but I could be totally wrong. When I mean high density I think about whole column is active in the same time.


#16

Hi @jhawkins, sorry for the delay in getting back on this.

The behavioural arguments for top-down activation of L2/3 are mainly related to visual illusions such as boundary and surface completion, where we “see” lines and shapes which are not there, but which are implied by distant features such as apparent corners; resolution of highly ambiguous or initially meaningless images until their content is described to us, and finally the evocation of apparently sensory experience by verbal description.

This paper provides evidence of L2/3 pyramidal activation by sustained injection of current to apical tufts in L1:


The authors suggest that this effect is not well-known because L2/3 apical dendrites are much thinner than most L5 apical dendrites, so the work is significantly trickier. In addition, I suspect the magnitude of each dendritic spike is relatively smaller in L2/3 than in L5, so it may take more of them to cause suprathreshold activity at the soma, and the effect will likely be rarer.

It’s plausible that the L1-L2/3 evocation of somatic spikes is only significant when the L4 input is absent or has low signal-to-noise (eg when you close your eyes), and that most of the time L1 only causes partial depolarisation.


#17

Hi @jhawkins
I am really impressed by your work. That is awesome. I started to learn HTM and read your papers recently. I have some questions, maybe they are basic questions, but I really appreciate if you guide me regarding them.

In " Why Neurons Have Thousands of Synapses, A Theory of Sequence Memory in Neocortex" paper in part 5. Materials and Methods you explained about the matrix of set of active cells A, and predictive cells \Pi, and a matrix of set of distal segments D^d_ij. This matrix D^d_ij shows all cells are connected to segment 1 when d=1, or all cells are connected to segment 2 when d=2, and so on? I mean what are exactly elements in matrix D? How we should consider segments when we are writing them in this format (by matrices)? zero, ones show junctions (synapses), am I right?

You mentioned “We assume that an inhibitory process has already selected a set of k columns that best match the current feed forward input pattern.” wuld you please explain me, what is the meaning of the best match? How we get it?

Thanks,


#18

Each segment is connected to a subset of the other cells. Typically this is a very very small subset. For example, in NuPIC, N=2048 and M=32, so NM=65536. A segment usually connects to only 20-30 other cells.

Thus D^d_ij when d=1 will be a very sparse matrix of size N X M components, with only 20-30 of the elements set to 1. The rest are all 0.

In a typical HTM system we 1) encode the data into a binary vector, 2) run the Spatial Pooler on the encoded vector to choose which minicolumns win and become active, and then, 3) run Temporal Memory on that set of minicolumns.

In the Thousands of Synapses paper we just start at 3). W^t is the output of the Spatial Pooler.


#19

Thank you very much @subutai for your response. I want to make sure that I understood it very well, so based on your answer, in matrix D we will show connections for each segment, that is right? so segment 1 (d=1) is common for all cells? then the number of these segments should be fixed?


#20

D^d_ij is a matrix representing the d’th segment on the i’th cell in minicolumn j. The segments are different for every cell.