How to Model Neurons (with Jeff Hawkins)

This will be focused on internal engineers, but Jeff is fine with us live-streaming it as well. Today in less than an hour.

12 Likes

Jeff is live now…

5 Likes

Absolutely brilliant video. Thanks Numenta!

15 min 28 seconds into the video Jeff talks about observing the spikes in his jaw muscles.

I had to try this! And indeed I heard it too. It also works when I frown my eyebrows. >:-).

1 Like

Thanks Matt and Numenta. This helps to reinforce my existing knowledge on pyramidal cells within the neo-cortex. As always it also leads to further questions :slight_smile: Such as: Are electro-magnetic effects at play wrt oligodendrocytes myelination of the axon (e.g. frequent axonal spiking attracts/encourages rate of myelination)? Are there inhibitory synapses found everywhere on the axon, or restricted to chandelier neurons affecting the axon hillock? Or wrt to myelin, inhibition that can occur at the nodes of ranvier?

1 Like

Hi Richard,
Myelination makes action potentials travel faster along the axon. I am not aware of any theories that the Myelin sheath plays an information processing role. If you know of any please share. We see Myelin as part of the “plumbing” required to make neurons work and therefore it hasn’t played a part in our theories. I suspect there is a body of literature related to Myelin as it is lost in some diseases.

As far as I know, there are no inhibitory synapses on a axon other than near to the cell body/axon hillock. There is an exception to this. In a few places in the brain there are inhibitory synapses formed on the end of an axon where it forms a synapse. For example, the triadic synapses on thalamic relay neurons, but this is not typical.

5 Likes

An excellent video.

For a newcomer to HTM, what really clarified things for me: Because a set of proximate synapses need to fire on the dendridic tree for a dendridic spike, I am now visualizing the dendridic tree to effectively have branches or sections associated with each of the linked patterns that a given neuron is responding too. The other “aha” moment was the notion that a dendridic spike enables a neuron to fire earlier, and that the inhibitory neuron that fires after the spike is actually shutting the other related neurons down, effectively declaring the fastest spiking neuron the winner.

6 Likes

Do you think inhibit neurons could play a role in eliminating unfitting scenarios in memory selection?
Like the firing stops when the prediction is different from reality :smiley:

The mini-column do what is call bursting and fire even more. This is well described in the Numenta papers; if you are interested in learning more see the BAMI paper here. Search for the keyword burst.

2 Likes

Given the discussion in this video, the concept of bursting described in BAMI (and other Numenta papers) now makes a bit more sense. In those papers, bursting is described as every neuron in the mini-column suddenly firing in response to proximal stimuli for which none of the neurons had recieved a predictive depolarization from the distal dendrites. If none of the neurons fired early (by being in a predictive state), then the inhibitory neurons did not receive an input spike ahead of time, and therefore were unable to inhibit any of the nearby pyramidal neurons. Hence, all of the nearby neurons were allowed to spike in resonse to their proximal inputs.

@Bitking Unless I miss my guess, it’s not that the cells are firing more often, it’s just that there are more cells firing simultaneously.

6 Likes

Yes, a better way to phrase it!

I think I also have a better handle on how the synapse permanences are updated by the neuron activation. Considering that the activated synapses have been recently flooded with ions, they could conceivably be considered to be slightly more conductive than the non-activated synapses. In which case the voltage spike (resulting from the firing of the neuron) travels back down the dendrites preferentially in the direction of the recent ion exchanges. This would sort of be like a static discharge following the path of least resistance to a more neutral state (i.e. whatever passes for ground state). This additional charge/voltage exchange arriving at the recently activated synpases might then allow the synapse to be strengthened through some kind of metabolic response.

The main take-away for me is a better understanding of the biological mechanism for how it is that only the most recently active synapses are reinforced.

5 Likes

I had a twitching muscle in my thumb the other day. I pressed it up to my ear and sure enough, I could hear individual spikes clear as day. Very cool.

2 Likes

Great talk Jeff.

Questions:
So to sum it up, one should consider a network of cells always rather than single cell like the SDRs?

With regards to old neuron model and biological one, the old one doesn’t take into account population affect?

I’m not exactly sure what you mean with this question, but meaning is only present within populations of cells. A certain SDR or sparse code has meaning, it means exactly something in your mind, and the meaning is not discrete, but can change subtly in so many ways over time.

If you take one neuron and ask what it means when it activates, it won’t know at all. It will only know that it fires certain ways in response to certain dendritic input. It is a part of thousands of bigger patterns playing out over time. If this cell dies, the system lives one as if nothing happened. The meaning is learned over time and is represented by the structure of the cellular networks that the SDRs express themselves over.

It certainly does. I think the old model (used in all of today’s Deep Learning architectures) is completely dependent on the statistics of population effects between layers of neurons. ANNs have the right idea, but the neuron is too simple. Both systems depend heavily on meaning being distributed, but DL does this in a very dense and computational way that I don’t think is happening the same way in our brains. By expanding the neuron model a bit to account for prediction, you can create sparse networks that work much more like brains.

1 Like

Blockquote
I’m not exactly sure what you mean with this question, but meaning is only present within populations of cells. A certain SDR or sparse code has meaning, it means exactly something in your mind, and the meaning is not discrete, but can change subtly in so many ways over time.

When looking at any task today, done by the artificial neural networks, all I understand is what a single neuron does (output I mean) and not the population of neurons (even though the final answer includes/ accounts for results of all individual neurons. But the way Numenta is looking at it is different in regards to the fact that they consider the many populations of neurons and the relationship between populations, if I am not wrong.

1 Like

Blockquote
I had a twitching muscle in my thumb the other day. I pressed it up to my ear and sure enough, I could hear individual spikes clear as day. Very cool.

Is there somewhere talked about in neuroscience field?
“When do decide to try or experience a task, you actually experience it, even though, all by yourself you have never experienced it ,until it was told?”