Numenta Live Streams, 1 Apr 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Numenta Research Meeting, April 1, 2019

Description

Numenta Research Meeting - neuroscience / artificial intelligence (AI) / neocortex oscillations https://gist.github.com/rhyolight/59dcd4f5810a00b001697abd70452411 -- Watch live at https://www.twitch.tv/rhyolight_

A

B

Good we're good to go.

A

A

B

You guys can hear me so it's my computer just crashed.

B

C

So it's gonna be a short meeting because subitize not here and the topic I have- is not very interesting. Well interested. So when when floor is here on Thursday, we reviewed this paper. The working memory, 2.0 paper and I was finishing his topic, and but it was built around these camera frequency members and other duties and the gamma frequencies for refreshing this temporary memory and then, but when reading this paper, I found a whole bunch of what look like very interesting references to other oscillatory things and so there's a whole.

C

You know because in my mind the oscillations have always been this sort of, like weird people studied them for like going back. You know 40 years ago because they found these oscillations. Just using me, you know scalp recordings, you know, but it was never like what are they for and so there's always been this one hypothesis that they could be just there to synchronize spikes so that neuron work better, but maybe they do something else too. So here he was with Lauren came in with a very concrete example of doing something.

A

C

Is using this, the gamma frequencies for recording short-term memories, I thought it very compelling, and then there was all these other papers and became clear that they've been a lot of research done on oscillations in the last 10-15 years, and I hadn't really looked at any of that stuff, and so they looked interesting. The title, which is really interesting, so I, went and printed a bunch of those papers and I read them this weekend. I read several of them.

C

I was a little disappointed over all that is, they sounded really primer trained in the ranging, but not as interesting as I thought they would be, and so here's here's one. This is a 2010 paper in neurobiology dynamics of active sensing, perceptual selection and even here in the introduction they talked about active sensing, which is you know, moving and there's going so much role of frequencies and oscillations in active sensing, but they say. Oh it's.

C

Obviously that's dissenting in, like the things like the bat you know which sends out its ping, but you know inconsistent contrast. Biological sensors, like the eyes and fingertips, have traditionally been thought of as passive sensors and I'm, like really even 2010. People are thinking, like you know, eyes and fingers its passive. So it's like okay, so they're, just like moving beyond that they're trying to show no. Those are really active. Sensors. Okay, this paper here this one's called the sensual cycles from cell press. This was 2016 paper by roofing ban ruling.

C

It was a review which is always nice and in this paper is called perceptual cycles. You get the first to get to the city I'll just share with you guys took this small. You memory come around. It's like this figure here figure. One he's basically, there's two sort of basic ideas of people are thinking about. One is on the left, which is a rhythmic perception, meaning like you're, representing something, but at the peak of the oscillations.

C

More neurons are firing than other times and therefore synchronize with them, but there's not really doing anything else other than that. That's just like the plumbing idea, like we're just trying to get the cells apart. At the same time, the author's idea is actually no that it's not like. Those neurons are just being getting the force, there's actually period of time when one one thing is being done, improve the time where something else might be done. It's like.

C

So it's not a continuous thanks to discrete perception, even though you might not be aware of it consciously, but the idea that there's little windows in which something is being calculated and other windows on the other part of the face or something that gives become later and the large part of this paper is making the argument it's more of a ladder that this is what's happening here and then, of course, the first time that at least I was aware that something would be calculating on different parts, and they mentioned that in here is the precession in the hippocampal complex, where you can show the different time frames.

C

Different cell populations are firing, so a lot of it is just sort of talking about the basics. These papers are like well.

A

What is it like? Those.

C

It like this and it's more like this, and maybe maybe there's other computations going on in the middle of all this, and these people are complex and hard to read. So, there's a lot more detail in here, but I'm not going to try to summarize and then.

C

This is this paper here. This is a 2010 paper, Bushman and Miller. Miller was on that working memory paper and it's called shifting the spotlight attention to evidence for discrete computations and cognition. So this is like going beyond the one. I just talked there with just saying, like maybe during that during attention that the tension is tied to these different oscillations and there might be different computations going on at the points and the attentional cycle, and one of the ideas there is that is, is like. Oh, what I'm attending to something the attentional aspect.

C

That's happened in one part of the cycle and then maybe I'm doing something that's not attended. On the other aspect. The cycle which we talked about is maybe a really cool idea for how we could do a displacement that you might actually be alternating between an intended object and an unattended object. Intended object, an unattended object from every single cycle, and therefore you could you could calculate you know. Over and over again you can be calculating the displacements as you're doing that and.

C

Yeah I'm trying to preserve again our cognitive, computations, discrete and they're, arguing that they are.

C

They're arguing that you can hold to multiple objects in memory. At the same time, the tended objects, an unattended object. So I'm not gonna, go through all the details of this, but in some sense, they're arguing for that. What we found in the hippocampus with the pre-session, something equivalent would be likely happening in the cortex and but.

A

C

They guess they don't get too far. I mean is what purpose might be discrete computation serve? You know we rename. The kind of you know sock. It's not like a very well form. Theory, I'm, leavin, I added more to it than they say here. It's basically well, you could be doing multiple things at the same time and they don't speculate what those might be, and they may also point out here that, although the proposing this is common throughout the New, York Court jacket, so you don't, they should also say it doesn't require all computations.

C

The brains. You'd be discrete like this. Some.

A

C

As great something may be discrete so I found.

A

That this particular.

C

Example they gave it's difficult to follow. It would require much closer reading and but I took I just accepted their conclusion. There's a conclusion is that we have these discrete time frames that we can calculate different things at different points of time and I thought that that is a whole new perspective that we've kind of been working away there by accepting the procession in the hippocampus and now there's a bunch of people arguing the same thing happening throughout the neocortex, even if they don't have to put theories for.

C

Why or what it's doing, and then this last paper I read here is a 2018 paper August that just came out and it's talking about spatial attention and and it's to see. If I can trying to be summarized again.

C

Here's one sentence: it says this is a this is by Viking coin. I know mining these people because it's a different field, dynamic interplay with the frontal parietal Network underlies rhythmic spatial attention, says converging. Evidence indicates environmental sampling, whether through covert or over mechanisms, as a fundamental rhythmic process, with a sampling rate and the theta band. So.

A

Again, more evidence that.

C

These these cycles are occurring between different parts of the neocortex and that attention is occurring on the cycles they have these bright. They have these very specific attentional paradigms that monkeys can do just a that. They test for and it's pretty damn complicated, but in the end, you're, basically just trying to show that many of these operations are tied to rhythm, that the secod initiate at a certain point, in a rhythmic cycle and in a whole bunch of things that are tied to these rhythms, but neither what a little disappointed was.

C

There wasn't like a more cogent theory about how tie all this together. That's that's the thing that was missing. It's.

A

More, like hey.

C

This stuff is happening, and these are active senses. There's these cycles going on the tension is tied to sizable movements are tied to these cycles. Maybe there's multiple things we represent at the same time during these cycles. This is happening in in your project and all the much more detailed parts of the paper seem to be not illuminating, deeper insights but more sort of supporting those conclusions. So I think it's it's interesting and it opens up the whole a series of possibilities for us and there's a ton of more literature out there. Everyone buried them.

C

Although I was reading some very recent papers and those recent papers didn't say how we figured this all out. Here's what's going on. You know it looks like no, it's a knitting area to study and lots of evidence suggesting it so that.

A

C

Guy's a little disappointed, but you know these are hard studies and people worked on for years now and and it certainly hasn't advanced beyond what I looked at the last time that, like this is 20 years ago. So so that's my report on those papers.

A

So general question I have on oscillations and I'm, not sure if this is known that mostly I've mostly only write about oscillations like by reading other things, that mention them and then I read a little bit, but I haven't dived headfirst into them. So what confuses me with them? A little bit is: are they a cause or are they an effect, because one way you could get an oscillation in local field potentials or you a noticeable oscillation is suppose there is a large population of cells.

A

That's firing at 40 Hertz, like suppose, suppose, a working memory is held by at certain times. At certain times periodically there are cells firing at 40 Hertz that could cause the oscillation that you detect I mean the. If you well I mean you're just measuring local field potential, yeah I'm. Definitely a large population of cells start spiking, it 40 Hertz, but why do they own it? The.

C

Question is why this I'm confused by commented Wireless I, mean they for them to spike together. There has to be some sort of underlying mechanism to make them spiked together.

A

C

So, are you suggesting that I.

A

Don't know it okay, explanation a so when you're holding in a memory or when you need to refresh our memory Allah to cause it to stay. You know when you need to cause a bunch of cells to burst in order to preserve a memory, your brain, somehow Ram cells, gamma yeah. It causes basically every cell to be receiving kind of the oscillating input that causes every cell to fire more often yeah and that's what gamma is something that the brain can ramp up and down. Yeah explanation be: when you some neural mechanism exist, they.

C

A

Cells to occasionally fire at about 40 Hertz and that can be detected in the local field potential as a gamma oscillations, yeah and I, don't know with like there's a.

C

Good well and I'm: switching up yeah yeah yeah! Well, you think they're firing if the firing synchronously at 40 Hertz, wouldn't it more indicate the former than the latter I mean probably yeah.

A

C

Think the attitude I mean interesting question: the attitude in these papers- I it comes across as they're all thinking the former. It's like the oscillations are a means by which the brain decides I'm, going to return to something. I now establish a an oscillatory frequency or I'm going to these two regions and your cortex are come talk to each other and that's the way they describe it.

C

It's interesting question: maybe it's more of an epoch bomb and that comes back for other reasons, I think that's what you're saying yeah yeah I think they thing for me, though you know all is I could, from an information point of view, it was never clear to me. Oscillations were essential if they're, really just providing this sort of synchrony I mean you know. Okay, we need to get the neurons fire together, that's part of plumbing, but the fact that they right now we've got these multiple things going on in the same phase.

C

That's a very interesting computational problem issue that introduces today. We've probably we were gonna work, the way they are ready. We in our sequence memory boards talked about hey. You know that we could be doing and inferring on the two cycles to do phases and that would solve a lot of problems. Oh so that was like. That was a big insight and now thinking about displays of themselves.

C

This could really play a big role in that kind of thing, but also we can start thinking about how do I get to representations alternating back and forth and that I can compute between them. Things like that, so that also when we start thinking about multiple things happening on the same octave phases is like wow. That's like a whole other information processing, realm I, don't know how that relates to your question.

C

Mark, as it seems means Susan means its oscillations that become this more fundamental information processing capability and therefore the brain is going to actively make them happen.

C

Somebody talked about that. You know there's some theories and they're just like. Oh okay, if you have an image or a cell and a positive and excited ourselves and hooked up a certain way to get oscillation, but mostly I, think these papers were just sort of trying to make an argument that that they are occurring in the neocortex and that they're there they mean they're, meaningful and related to things that we know that animals.

C

Certainly animals can do yeah, yeah, there's a whole slew of litter dramas and again, my reading of it today and this weekend, sort of implies that they spent a lot of time. Reading and I don't want to get the sliding for mental knowledge. Young boy, if I'm reading a person like the August last year, and it's still talking about sort of basics- it's like! Oh, it's moving into their projects, but it's not really super fun. Super well understood yet, and.

A

I'm guessing a lot of this is just the fact that our measurement tools we have oscillations are really convenient thing to measure you just like you're, using EEG or any of these types of these types of recording tools. And then you just what perform a Fourier decomposition and like there. You have the signal, so it's kind of an easy thing to measure, and that would cause.

C

What they're trying to do now, though, to understand all this other stuff, so you have to look at single cell or multi cell report. It yeah, because if you look, it didn't phase it like. What's the population code at this point what's population at this point, so all of them, these studies, I think I. Think I should go back. Nobody all over or most some were like monkey macaque multi in a recording and maybe similar as human I can multi unit.

A

C

And you know getting more underneath the covers of it like, oh, because because it's really helped me think about it. You.

A

C

This this EEG recording, which is a reporter with 15 or whatever.

A

C

A

But it really could.

C

Be individual subpopulations off at different times, like yeah I thought Florida's presentation last week was pretty convincing. You know, like I, think, like they really did nailed down this short-term memory thing at least now, like that's enough. That's pretty impressive anyway,.

B

C

The homeowner we've been working away in this area anyway, we've been trying to think about how these incidents our models and I just think. It's more evidence. Okay, this is another tool in the brains, and the cortex is a toolkit for representation and calculations between representations.

C

I'm not gonna, read anymore these papers at this point in time, less no, it's good. It was diminishing returns. That's all I got for today. Okay else,.

A

I, don't have any all right max.

C

Nope all right, we're gonna call it quits for there.

C

Yeah I'll take it back.

B

Okay, thank you.

B

It's a long react.

A

B

Goes: okay thanks.

B

A

Matt you're still on the go at the meeting I'm going to turn the dock.

B

Though I was trying to get off thanks, okay, close all right there, you go so that's that's a research meeting that was pretty quick, they're, usually a little bit longer than that. But it really depends. It's all you guys can hear me. Okay, it really depends on who's, got topics and who wants talk about what? Sometimes these things will go on two hours, they'll come back after lunch and continue it or it'll go into lunch, so it really just depends. So this was just like a one sample and to them into research meetings and I'll.

B

Just keep streaming them and give you guys some context into. What's going on at Numenta sound as low I know why this sounds.

C

Slow how's that that's better.

B

So thanks for joining it, does anybody have any questions about what you just saw I'm happy to stick around and answer questions is just. What's spell it right? If this helps anybody, we just are done with the meeting. Now it was, they were talking about oscillations in neocortex. There wasn't a whole lot of mention of thalamus six. It was a it was sort of assumed and I thought. It was interesting. The idea of our oscillations cause or effect I think they have to be both right.

B

The oscillations have to be caused by something, and then they probably cause other things. I mean that's it's a part of the compute process in the brain, so so yeah I, think I. Think the general idea is that gamma oscillations in the neocortex are are suspected to be caused by the thalamus. Although I mean there's still a lot to be learned, the hard thing is experimenting.

B

It's hard to find willing subjects, because you know you have to open up their brains and put probes right down in there, and you know that there's ethical concerns with that.

B

So we try and do as much as we can with fMRI, which is an unobtrusive way of monitoring neurons, but it's very hard to tell the connections the directions of the connections and the it's, because there's there's breakthroughs happening right now, so I can't really speak to what the breakthroughs are, but typically it's a measure of blood flow and what we really want is a measure of voltage and it's very, very particularly localized areas of the brain.

B

So we can, you know, focus on specific areas and what those neurons populations are doing so anyway, you that's it so I'm, gonna, log off and I've got another meeting.

B

Well, I'll tell you what, since we're since we're on twitch we'll do we'll go raid. Somebody since I got some some.

B

I got some followers. If you guys want to read somebody: let's do it. Let's see who's who's streaming right now, who haven't I raided, yet this code enjoy guy. That was this. Is it's this guy's birthday we could read it. Let's do code enjoy. I watch. I watched him a little bit the other day. Oh he's playing a game: I'm, not I. Don't usually raid games who's, this guy CG CG games, oh yeah I was watching him earlier I enjoyed that I say: let's trade him!

B

Okay, here we go, didn't see any more questions, so we're gonna get on with this and I I think that work did that work.

B

There should be a yeah, so we're gonna raid. This guy. Thank you for watching I'll, be back at 1 o'clock today with an AI chat. We're gonna talk about pooling, different types of pooling, pooling, like in CNN's and pooling in space, like spatial pooling temporal pooling those ideas too. So, let's.

C