Why Does the Neocortex Have Layers and Columns, Learning the World's 3D Structure | Hacker News Reader

Why Does the Neocortex Have Layers and Columns, Learning the World's 3D Structure

biorxiv.org

109 pointsjcua35 comments

Former neuroscientist here.

I have no idea why this paper is at the top of HN, other than that the former founder of Palm is the lead author.

Setting aside that it hasn't been published or accepted, the only remotely novel thing seems to be they suggest that location information is represented in each column, and this occurs everywhere in the neocortex.

Well, what does the actual biology do? On the one hand, it's well-known that the trillions of long-range connections mean you can find neurons almost everywhere that will fire in response to place/location stimuli.

On the other hand, the entire medical history of restricted lesions and targeted deficits (Broca's area, Wernicke's area, patient HM's hippocampal anterograde amnesia, visual scotomas, etc etc etc) indicate some brain areas really are specialized for some things more than others.

The brain has some redundancy, but it's such an expensive organ (2% of body weight, but 20% of your energy budget) that I would not bet on such widespread replication if avoidable.

> On the other hand, the entire medical history of restricted lesions and targeted deficits (Broca's area, Wernicke's area, patient HM's hippocampal anterograde amnesia, visual scotomas, etc etc etc) indicate some brain areas really are specialized for some things more than others.

Wouldn't the evolutionary history of the brain imply that information is encoded in two different ways in a lot of cases. Ie. that the components of the older 'lizard brain' (if you'll excuse the issues with that term) have a rather fixed function, but the neocortex evolved overrides to those functions?

Like sure, we've got a much better understanding of grid cells and place cells in the hippocampus, but does that mean that there isn't similar information being stored in a different way in the neocortex?

but the neocortex evolved overrides to those functions?

In ideal circumstances that might be true but in fact it's usually the other way around. Kahneman's type 1 and type 2 thinking theory expresses this most succinctly:

The key theme of the book is that the human mind has two distinct systems of thought. The first, which Kahneman labels System 1, is fast, reflexive, intuitive and automatic. It is the primitive part of thinking that evolved to allow us to survive in a dangerous world.

System 2 is slow, rational and deliberative. It is the part of our thinking that we can consciously “observe” and is used for analysis, logical reasoning and deliberate calculations. System 2 can sometimes act as rational governor, overriding judgements or decisions made by System 1. However, the main takeaway from Kahnemen’s life work is that System 2 can be “lazy”, allowing System 1 to lead us to irrational conclusions.

In some ways that's true, though the more recent areas tend to cooperate/elaborate, not replace.

E.g., both the superior colliculus in the midbrain and the frontal eye fields in neocortex are responsible for making eye movements, but the FEF doesn't override the SC. Even in this case, it's not like eye gaze/movement info is dispersed throughout all the cortex, it's limited to a short list.

We also know that the receptive fields (which part of a scene a neuron fires in response to) increase in size as you go up the visual hierarchy. V1 neurons will response to tiny parts of a scene, whereas V4/MT and V5 will response to huge swathes of your vision, indicating they become less location-sensitive. Object-identifying neurons may well respond to the object anywhere in your vision, suggesting they're not location-sensitive at all.

I guess Hawkins might argue that it's different neurons in association cortex columns that encode location, but frankly, the argument from biology is poor. There are cortical areas mapped out like this... but I can't imagine why Hawkins thinks this is a governing principle for the whole cerebrum.

Thanks for the elaboration! May I ask why you're a former neuroscientist?

Biology is messy. My data weren't so hot. I didn't want to live in small college towns at the time. Had more fun writing Clojure side projects.

Many, many reasons.

This video gives a little more context: https://youtu.be/fhnMUc36opI

Question - do you see any potential that horizontal versus vertical (with respect to system, not just euclidian space) has to do with cognitive stacking ? For example, is it low level feature detectors in initial visual processing that cause us to jump because they fire a 'snake' stimuli and then higher levels of the cognitive stack catch up and fire 'sorry, my bad' ?

Maybe. I don't know as much about cortical layer differences, but stuff like you describe (fast threat processing) tends to use different, faster pathways through the brain.

Intuitively, this makes sense, because the connection time between two different layers in the same brain tissue is miniscule, but the connection time between different brain regions is much larger, and is therefore likely to dominate overall response time.

For some context, this paper is from Numenta, Jeff Hawkins' company. Hawkins was the founder of Palm/Handspring back in the 90s. Afterwards he became quite interested in neuroscience and neural networks, and pulled a Wolfram in his 2004 book "On Intelligence" with claims that he had a unified theory of how the brain works.

https://en.wikipedia.org/wiki/Jeff_Hawkins

As a former neuroscientist, the majority of people who claim to have grand unified theories of the brain are those who haven't studied it enough to know how far we still have to go.

Well, Hawkins isn't one of them.

neuronexmachina's flippant comment misrepresents Hawkins and On Intelligence, which lays out in popular form a scientific theory of neocortical function, with perfectly sane perspective on what that means in the context of the whole brain, due references to neuroscience, suggestions for experimental tests and obvious awareness that the theory may prove wrong or incomplete.

I'm disturbed to see that two of the early comments here contribute nothing of value and amount to little more that personal attacks by way of misinformation.

Let me clarify. I'm not just a former neuroscientist; my field was consciousness and perception. Rightly or wrongly (as you may believe), nobody in the field is reading Hawkins.

What we do see a lot of, are people getting enamored of consciousness and dipping their toes in. For all I know Hawkins is an unheard-of genius, but his stuff reads like many of the dilettantes I've already seen.

Nevertheless I appreciate his efforts and just bought his book, because having an idea is way different from putting own money in it. He puts his own money to neuroscience, and does not compete with real scientists for government grants.

Other millionaires may buy yachts, or just spread "novel" ideas (e.g. Hyperloop idea by St. Elon) without investing their own money.

I appreciate and agree with your perspective on putting your money where your mouth is, but Elon is the last person to criticize in that context. He quite literally put all of his money into the problems that he thinks are the most valuable to work on (whether you agree with him or not).

False. He didn't put a single dollar into Hyperloop (idea that was researched since 100 years ago, and general conclusion is that it's not worth it, too expensive to maintain).

The book being from 2004 is kind of outdated. Nevertheless fun to read.

Agreed, hence my comparison to Wolfram. Like Wolfram's "A New Kind of Science," he wrote a nice book that was a pretty decent layman's summary of the field, but includes conclusions that aren't anywhere near as revolutionary as he thinks they are.

Hawkings was the PI at the Redwood Neuroscience Institute. Which is now part of Berkeley. He is not an outsider (15+ years of research). You could argue though, that Numenta is more focused on bringing these methods to market, then solely doing research. I get that this doesn't always resonate well with people in academia.

Are you sure he was actually a PI, and not just a funder? According to their list of publications (http://redwood.berkeley.edu/wiki/Publications), Hawkins is not an author on any of them. He's also not listed on the "People" page (http://redwood.berkeley.edu/wiki/People_at_the_Redwood_Cente...), even as an alumnus.

Question from someone who is not knowledgeable in ML or AI: Do current implementations of ANNs allow trying out these different types of organizations/structures? Is the structure and workings of neurons sufficiently similar to ANNs that the architecture transfers, or are there huge differences that make understanding and development orthogonal?

Most ANNs are not similar enough to biological NNs to attempt to recreate these structures.

A feed forward network is just matrix multiplication with a nonlinearity. It could, at best, be described as biologically inspired.

Also SNNs seem to be an active field of research. Namely research center Jülich (human brain project) is conducting experiments on modeling & simulating biological neuron populations.

https://en.wikipedia.org/wiki/Spiking_neural_network

http://www.nest-simulator.org

The short answer is no, the architectures don't transfer. At least not today. Biological neurons are much more complex, and therefore behave quite differently from neurons in neural networks.

To put things in perspective, we've fully mapped the connectome (map of neuron connections) of the simplest animal, C. Elegans, which has only about 300 neurons, yet we still can't simulate this organism's behavior computationally.

Last I knew, OpenWorm was making good progress on the nervous system simulations, though. IIRC they’ve demonstrated swimming, retracting when bumping into walls, and food-seeking.

Feed-forward CNNs pretty much describe V1 to V5 in primates. There are multiple papers on these architectures in the last decade.

And another question on top of that: is there any biological theory that corresponds to backpropagation?

A really interesting question. It would seem far fetched at first that backpropagation could be used by the brain, because it's unclear what the mechanism for transmitting the error at every synapse backwards, to the synapses in the previous layer, would look like physically.

This is a really interesting lecture given by Geoffrey Hinton (https://www.youtube.com/watch?v=VIRCybGgHts) where he discusses the various issue commonly raised with "biological backpropagation" and proposes a solution based on Spike Timing Dependent Plasticity (STDP). Basically he argues that you can interpret the STDP learning rule as a derivative filter on a firing rate and get backpropagation in this way. This is just on wild idea though and has not been shown to work experimentally or through simulations.

There are also a couple of interesting pointers in this Stackexchange thread: https://cogsci.stackexchange.com/questions/16269/is-back-pro...

You should check the neuromorphic community, really interesting work/people. They go from analog electronics to diehard neuroscience. I loved their Telluride summer school.

E.g. http://journal.frontiersin.org/article/10.3389/fnins.2016.00...

I once attended some talk or other and that question came up - the answer given was flippant, but I always thought it was oddly perceptive...... "Dreaming"

So the implication in the title is almost certainly false. There are columns and layers in areas of cortex that are responsible for processing sensory modalities that are not inherently 3d in nature. If I had to guess (from my vague recollection of the literature) the current thinking is that laminar structures are a nice way to build things developmentally, and columns arise in part from development but also because of *otopic mapping, which is to say that many sensory modalities have a reasonable mapping into two dimensions that have correlations in their inputs based on the destance their sensors are from each other. Obviously not the whole story, but laminar and columnar structures are very likely not 'just' about building 3d representations.

All that said, I could buy an argument that cortex has settled on a 3d representation because, even though the exact encoding of muscle movements is most certainly not in x,y,z or other 3 vector representations, a large portion of the motor actions that we make are indeed in 3d space (a good counter example being language which has representations in cortex and acts on a very high dimensional space of our vocal and facial muscles).

I interpreted the headline as though they meant that the neurons were in a 3d structure and part of the world.

He isn't claiming that this is how the brain works, he is just proposing a model. He describes in some of his talks that his work isn't about replicating the brain, it's about distilling algorithms from brain function/anatomy.

There is a short supplementary video, by the way:

https://www.youtube.com/watch?v=fhnMUc36opI

This paper makes use of hierarchical temporal memory.

I recommend reading this discussion: https://cs.stackexchange.com/questions/13089/some-criticisms...

Reply on news.ycombinator.com