Numenta Office Hours, 8 Sep 2014

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: NuPIC Office Hour - Sep 8, 2014

Description

Topics: https://github.com/numenta/nupic/wiki/NuPIC-Office-Hour-Sep-2014

A

Hello newpick welcome to the dupek office hour uh this. I don't know what this fourth or fifth one. um This is uh september, 8th, coming up september, 8 2014.. We have.

A

So this is a large part of the engineering team and we're here to do some uh some q, a um my name, is matt taylor, I'm behind the camera. So I won't bother you with my image.

A

So, let's get started, we we started off with some uh asking the mailing list, what some for some of their questions and since uh we've got uh expert online right now, who just recently participated in something called the world port hackathon and reported to the mailing list uh that one most innovative hack isn't that right, edward.

B

Yes, that's right.

A

B

A

Of the questions that you were referring to uh to the guys.

B

Here, you're ready for that now, yeah! That's fine! So you.

C

Could explain in a little bit what you did as well? That would be great yeah.

B

Okay, so uh we had some open data from um some companies uh they provided, and we tried to uh predict anomaly in incoming and current traffic in the port of rotterdam and and the environment and uh yeah we tried to see if newpic could predict uh some abnormal behavior and uh yeah. That was the hack we did in 24 hours so.

B

And uh one of the questions I run into was um uh we: we needed a lot more data to uh learn new pick, but how to handle the false positives. So um in in in like a big set of data for, for example, a year there may be some anomalies they wanted to detect in the future. But maybe newpick will see them as uh something normal, because it has already happened in the past. So how to handle that with, if you're, taking into account large volumes of data.

D

That would be more of a false negative right.

E

Well, the topic that we've.

D

Discussed quite a lot about you, there's um there's some ways to think about anomalies. One is that you're trying to classify them and recognize them, as if you know.

F

What they are- and you want.

D

To see if this occurs again, um that's really a classification problem uh and the other one is to basically.

F

D

Things that that are unknown, something that are truly different than you haven't seen before it's kind of hard to reconcile those two they're, really sort of different problems: one's a classification problem and one's an anomaly detection problem. So, even though some people use the term anomaly.

F

To say, well, here's an anomaly: I want to see how.

D

To see if it would tell me when it happens again, we really don't do that as an online problem, because the classification problem, uh we did a lot of work with with new pic and the html algorithms on classification, and uh you can do that. It's uh it's not what's in the code today, it's not.

F

Part of it normally a.

D

Detection framework, um and so that would be sort of a different effort. uh What we've we've found is, uh what we came to believe was that uh the real value most people, many people, come in thinking. Oh well, here's my anomaly. Tell me: when it happens, the real value is when you can tell things that they didn't anticipate as being anonymous, and so it takes a little bit of a frame of shift to framework thinking about that uh before people start seeing, that's really the outcome that's actually harder to get to.

D

I don't know if I can add any more to that anyway,.

G

We did have it on the classifier at one point I don't know, and I think there may be some code in it for that, but we haven't used it for quite some time. Yeah. I don't.

C

Think there's an example for.

G

C

After being on the classifier, but it should still be working because the tests are still running yeah, but it does require an elimination.

D

C

Yeah that does require.

F

C

Label data sets uh like like jeff, mentioned: it's really a supervised learning problem, so you have to say you know between here, and here is the pattern that if it happens again.

B

I want to detect.

C

Again so someone.

B

Has to go through and.

C

Label these things ahead of time.

B

Okay, um another problem: we are the problem uh I think we run into. If you know we were trekking around uh 4 000 different ships in front of the coast and they all have like continuous tracks. So so, if you look at the gps data and how we put it into a new pic, it works with tracks.

B

So like a starting point and an end point, but yeah we had problems defining when is the starting point and when is an endpoint for a ship and how to manage all that correctly and feed that correctly into new pick in order to get it to work properly. So what are your ideas on that.

F

So you're trapped, you're tracking.

D

uh Ships, large ships right.

F

D

And are you tracking them in a particular port or over larger areas.

B

Like um across the netherlands, so in the netherlands it's a lot of water, also inland. We have a lot of sluice for raising water and etc. So so we have a lot of ships in the netherlands and also in india.

D

I don't know what your data looks like, but um I have an.

H

D

On my cell phone that tracks ships and they typically they they have the last protocol and the next protocol, and so that kind of data would be pretty obvious when the ship leaves port. Wherever that is, you could start a track and then, wherever it gets the port, then that would be the end of the trip. If that was all contained within the netherlands, that might be sufficient.

D

If these ships are coming from elsewhere around the world or going elsewhere in the world, you might need to provide an end and beginning of those tracks um somewhere at the boundary of your border of those ones.

G

So that's also more of an application problem, so in the new big model.

I

G

Doesn't really care where the start and end is as long as you're feeding your data continuously for the same ship? uh You don't want to interleave multiple ships.

D

Yes, but I think maybe I'm just ready too much into it, but if you just track, if you just take the data from any ship forever- and you don't say well, here's the beginning and then it'll just it just won't. It won't it'll learn like one long high level sequence as opposed to like you, wouldn't be able to repeat it over and over again I mean if it's such a that just continually moves on for its entire life and it may not repeat itself at all for a very long period of time.

D

Yes, um but I think if you just said any ship, that's leaving this port uh and and you create a trap until you get to the next port if it was all.

F

D

Within the netherlands that might be a reasonably defined set of problems. uh I think you get into trouble if you want to track it out outside of the local waters of metal, and then you have to put an end to it. I'm not sure if you, if I understood your.

C

Comment, I guess there's the other issue, which is you don't want to so you want to give it coherent sequences of beginning and support, but you don't want to intermix shifts, so you don't want to give five minutes.

F

Of one ship five minutes of another ship.

C

And then go back to the first ship, that's what you were gonna do.

G

That's the that's. The only thing that I can really see is you're more just talking about. How do you select.

D

If you just continue to take wherever the ship goes, when you never say it arrived or left, then you would end up with this very long sequence. As the ship goes around the world and then we can take for training.

C

Is that what you were saying? Well, maybe you can tell us how you did do it and maybe a little more specific question.

B

Yeah, so we thought about two ways we could classify. The ship is, if we could say it's a container vessel, for example, and and use other ships that also contain a vessel to make sure we have multiple sequence. So we can see if there's anomaly in in the behavior or we could uh track every ship separately, but that could imply some ships arrive every five years or every seven years. So you have like like two tracks in 14 years, so you can't really do anything with that.

D

Yeah, I think the first strategy you set for example, sounds pretty good yeah. I would have.

I

Even gone further, though,.

D

I would have said, don't even classify the ships just say: a ship leaves some port and it arrives in some other port and obviously.

J

There's certain types of.

K

Ships will not.

D

Go to certain ports, but the simplest approach would be to not even try to distinguish between them. You just say you're, building a model that ships any ship and you have a beginning and an end to that trip, um and so if the ship leaves the port.

F

It might go a little container. Ship may go one way and a.

D

Bulk uh bolt carrier might go another way, but uh but you know first level you can just say I'll, stop either one of those, but if someone goes some crazy way, you've never seen before it'll flag it.

D

H

We thought about this just a little bit.

F

We thought about this in terms.

D

Of like airplanes might be doing the same thing and you say well, you might want to know the size of the airplane or the carrier, but you could just say no.

F

When an airplane leaves this airport.

D

um And it follows the track and then some other airport just assume they're all the same and then it'll quickly settle in on which, which is like you know it might say, well twenty different ways. It goes initially, but very quickly, they'll settle in on the path. I'm sorry, I noticed it was different way. You proposed it takes more data because I have to have more data for each type of ship and it's even worse. You track each ship.

D

So the quickest way to train the system would just say all ships are equal and just say you know just track them on the support.

B

Okay, another thing I I found very interesting was: maybe we could a ship has a course and heading where they're going, and maybe it would also be interesting if you uh uh combine the the behavior of also predicting where something is going and tracking uh anomaly. That was just a thought.

B

Maybe we could expand the to do that as well or or combine those two, but, and another thing was also- you talked a little bit about it earlier about uh marking certain patterns to see. If, if this certain pattern occurs, we want to.

B

See it, for example, I was thinking about um for for the coast of somalia. You have a lot of piracy going on and they have a hard time, detecting uh piracy ships uh trying to- uh and it would be very interesting if you could uh classify some kind of behavior and say this is a pirate ship. If anything on the radar uh uh makes us similar, uh yeah uh show similar patterns, then that would be an interesting vessel to uh look further into.

B

These are some of the things uh yeah.

D

That's that's right, and a lot of people would like to do that, but you have to remember that the flip.

A

Side of that issue start classifying and saying this is what a pirate ship looks like.

D

C

I

Just start behaving differently.

D

And then it doesn't.

F

Look like that anymore, the.

D

F

Thing about an online learning system is.

D

That it just detects when things change historically, but but that's really a different problem. If you're trying to classify something, it's not really an anomaly detection, it's a classification problem and you can do that. Yeah.

C

Going back to the earlier question, I thought.

K

F

Was very interesting: we haven't really done a prediction.

C

With geospatial coordinates, but I think you.

F

Can get some really interesting results with that.

C

So if you know that, um if your prediction.

F

Is again in the using the same in the same language as the encoder.

C

Because the cla.

F

Or the htm learning algorithm can make.

C

Multiple predictions, at a time you could say: okay, if uh chips start going.

F

In this way, here's kind of the cloud of locations it could go in, but you know it's not going to go.

C

In any of these locations and then the further.

F

And further along the course they go. The more.

D

You can fine-tune.

C

Those predictions I think that could be really cool.

D

C

A lot of work in that we know that it works.

D

Well, so the biggest problem we have with this is trying to figure out what to do with that prediction.

B

Yeah I so I thought that may be interesting uh if you think about fencing, maybe like an offshore oiling platform and- and you want to be notified if a ship uh is heading in that direction or the prediction is that the ship will uh go in that direction, then before it actually happened, you can address it, and- and that would be very interesting, I think so. It's not yeah.

C

J

A great application that would be really cool that should work well. It would require uh some extension to the geospatial inhibitor yeah with it yeah, because it would make predictions in the sdrs and then we have to map the sdrs back into the you'd have to do sort of reversion code information, yeah.

D

Well, that's a.

J

One-Way hash, so you would have to pick an arbitrary region and go through, but that's.

D

Not the way we did prediction, though right, uh that's like the reconstruction, but we did prediction by putting a separate classifier uh on the.

I

Internal state, you know the classified classification required.

G

That you see the points before where the geospatial, where you have a very high grade resolution, you might not see them right in the way that change is scrapping, doesn't matter. If you see them or not, it will still.

J

How did you do so? I don't care why, but you would predict the classifier would predict things that it's already seen before.

C

K

It would just predict things in xyz space.

C

Xy and speed that would be uh just.

F

C

Of a scalar value, we're just.

D

Yeah, so you basically classify the location um you use a classifier to legal location. You have to free how to label it, but uh in the classifier that.

A

Let me just show people what we're talking about real quick, um so I am.

F

B

A

Image of the I think these are both tracks within the part, uh the partner of rotterdam.

B

This is amsterdam actually but yeah. It's a port as well, so right.

A

uh Just so for people watching, this is what this is um edward's hack from uh past week after week, um and these are these are the geospatial tracks we're talking about, and one thing I noticed when I saw this- is that pretty much everything seems to be pretty highly anomalous, with just the with the cursory glance anyway. Well, you know, I said this train. It hasn't been trained much here, right.

B

Yeah yeah yeah, it's correct, so uh we um had uh some hard times with the api because we did too many requests around two at night and the server went down. So we only uh had like about 24 hours of data, but that was the problem. So.

D

I look so there's behaving well, it's basically saying I've got. You know, I'm tracking making these tracks. They haven't seen any of these things repeat enough to to, uh but presumably, if you did this for a month or so you start seeing these patterns become very well known. This is a great example. I think this would work very nicely yeah. It's really cool.

B

Okay, so that's one of my questions so yeah.

C

Right it'd be cool to keep going. I don't know if you're interested in this, but um now that you.

F

Have the data or maybe collect a little more data.

C

Work on this a little bit more, I think, that'll be a very cool demo in an application. Yeah.

B

Yeah I'll see what I can do.

A

Thanks a lot edward and congratulations.

H

On winning that prize.

A

Of that hackathon, it's great to see that yeah. Thank you, okay, so.

H

A

We don't have any uh questions come through yet so, if anybody's.

F

Watching this, this is now. This is when you should be.

A

Submitting questions, uh but we do have some uh questions on a wiki page that we gathered from.

F

The mailing list, so let me.

A

uh Try to figure out how to use my computer and share with you this page of questions here. um Let's not talk about pattern stability. Yet let's talk about the motor sensory work, so three or four people said they wanted to see. uh They wanted to hear more about.

F

What's going into this new.

A

Newpick resource repository so maybe stupid. You can explain why we're doing this and what the protocol is going to be for this new new research repository.

C

Yeah, so this is something where it's an experiment that we're doing um just to see how it will work. um You know we so we've released.

F

C

And we have, you know, that's commercial quality code that we use.

F

In our product and.

C

The new big open source community is maintaining it and we're maintaining it. That's an active development on top of.

F

That we're also experimenting.

C

A lot with new research ideas along the lines of sensory motor integration.

D

C

Building up the whole particle framework, we're always experimenting with new ideas and things. So this.

H

C

Kind of an experiment.

E

C

Can we give the open source community kind of a view into what we're doing there? um You know why? Why hide it? Why not be more open about it and share it? So um that's what we're that's the goal of of the new pick research repository.

C

We have experimental code that we use to try out all of those algorithm variations and stuff, and it's constantly changing and so on. So the goal.

F

Of new pick research is, we will just put our code into that repository day.

C

F

C

um And right now, there's.

F

Nothing there, except for a readme which pretty.

C

Much says what I just said, but so you can read that the other thing about the repository is that you know it's extremely experimental. It's very raw. It may be very confusing to people it's gonna be buggy. We may not have time to explain things, um so um you know it's sort of buy everywhere. It's it gives you some sense of what we're doing within a dementor, but.

F

It might also make you confused if you.

C

Think you know you might because we're going to be changing it uh constantly or.

F

It might not change for a.

C

While for a long time, for whatever reason, so that's a basic.

D

C

The idea behind the thing.

D

I just might add that the reason this came about is that you know we'll have an idea like how this how the sensory motor uh integration inference works, and you know we talked about that back in january february, or something like that, and then it takes a long time before we actually implement it and test it in various ways, make sure we go so people are saying well what's happening. You guys talked about this doing anything with it. Yeah we're working on it, but we don't. We don't feel.

D

Like you know, every day we can't explain everything, we're.

F

Doing every day.

D

Too much yeah, so uh this was a response to the request to people saying you're, really working on this yeah.

G

Here it is, you can see.

D

What we're doing, but without really we don't really have the ability to explain it at all or support it or anything like that.

C

So I can talk a little bit about the a little bit of the code that we're planning to put up there, which is our sensory motor research thing. So there.

A

Was an email thread on this and.

C

I think I I put in there some of the stuff that we're working on, but the basic idea is that we're working on the sensory.

F

Motor imprints and the.

C

Temple pooling ideas that jeff outlined in an email months ago- and there are basically two pieces to it- one is the sensory motor inference, so there we're modeling layer four of the cortex and the idea is you have sensory patterns coming.

F

C

You have motor commands coming in and the goal is to be able to predict the impact of making this particular motor command.

F

C

Sensory pattern, can you make a good prediction of that? So imagine you have like.

F

A static world or a sounding image.

C

And you're circuiting.

F

Around it or you're moving around this world.

C

As they move around given where you are, and you learn to predict where, where you will what you will see that if you succot over there or if you move over there, so that's kind of the goal of the sensory motor inference portion and then on top of that, we want to build.

F

Stable representations of this.

C

World, so that's where the temple pooling comes in, but given a particular world once you've learned how to uh predict what will happen.

F

C

You know, can we create.

F

C

F

Stable representation.

C

Of that whole static world, so you know in images.

F

And sakaz is another.

C

uh Maybe an easier example: you look at and a complex image and.

F

You're cicada around it.

C

Once you kind of know what that image represents, you can build up a stable representation of that image and that, well that's what would happen in layer 3 in the temporal pooling area. So we have that's kind of the two tasks.

F

We're trying to.

C

You know code up and.

F

Understand well so we have initial implementations.

C

Of that code, which will go up there and I can almost guarantee.

F

C

You guys will be confused when you look at it. The terminology.

F

Is all over the place? It's.

C

Messy code uh not well tested, um but the kinds of things we're experimenting with are implementing the basic mechanisms for those operations. Understanding.

F

The capacity limitations.

C

How many stable.

F

Representations.

C

Can you make how.

F

C

An object: can you understand how large a stable pattern.

F

Can you go through what happens.

C

With different types of motor commands, what happens if two two static objects.

F

Share parts of their representation.

C

What happened, if you uh exploit, you can only explore part of it.

F

And then you come back and you explore the rest of it.

C

Can you still build up a single.

F

Stable representation of.

C

That world, so those.

F

Are the types of things that we're.

C

Going to be working on and have been working on, so that's.

F

When we put the code.

D

H

D

C

um You know you'll see.

D

Experiments around a little.

C

D

Is the testbed where you can you can start off with a very, very simple world? That's, like you know a simple maze, but you can. It can gradually get harder and harder until the point you're trying to recognize an image or something like that. So it's a test bed that takes us through a lot of different possible problems, but it might be hard to describe yeah. Unfortunately,.

C

And it's a very difficult, I mean it's a pretty difficult and cool thing that this system will be doing. It's basically, you know so psychologically. You know, as.

F

We're moving around this world.

C

You know the actual patterns that are coming.

F

Into our senses are constantly changing.

C

F

Psychological perception is that of.

C

A stable or a stable image, but that's not at all reflective of the patterns that are coming in. So how can you build up a stable representation of a stable.

F

World, given that your inputs are actually constantly.

C

Changing and how do you learn about complex environments from that? That's kind of the problem we're trying to attack.

D

I believe this, this basic problem was recognized back in the 1800s, and the chemicals wrote about it um and it's been sort of every once a while philosophers come up this like how does the world seem stable because we're moving and everything's moving all the time our bodies are moving our bodies, but yet it seems stable uh and, of course, if you, if the world.

F

Actually does move on you.

D

And you don't have you know, then you get sick instantly. You know if the pattern shifts around in your eyes. You do not because you're moving your eyes that tells you your body is compensating. It knows it knows that, because of the movements that it's generating, it knows how to not make it sick. It makes everything look stable. So this is a whole problem, but not many neuroscientists or computer learning. Experts have thought about it, but it's something that's been recognized for a long time. Besides some smart people a long time ago.

A

Well, it's exciting that we're gonna! We have this new repository and over the next few weeks, we'll be putting more and more code in there and you'll. See that start changing and evolving as research continues: yeah.

A

um Okay, so uh video has follow-up questions about new pic research uh use the q, a interface I'm gonna, move on to.

F

Our old buddy rick.

A

B

A

B

Not nagging at all, but I have a question.

A

For suicide about the quantitative model, whether there's any interesting progress, uh what.

C

Type of this is the theoretical paper on str properties.

A

I

C

I've been on all the sensory motor stuff, so there has not been a lot of progress in the last few weeks. Unfortunately, I know I.

I

C

Need to do that so I'm.

I

D

It's really pretty good.

D

I feel like the.

A

Yeah all right um and let's move on to this- let's talk about the swarming.

H

Okay, I think these are some.

F

Interesting questions.

A

F

The mailing listings.

A

I hope I don't uh get this wrong. Rhiann shams, uh just mailed. I think this morning, this question some questions about swarming and why did we choose the pso algorithm versus other evolutionary, optimization techniques?

A

So, let's tackle the first question, which is what was the reasoning for choosing pso as a technique to determine the model parameters as opposed to other optimization techniques such as genetic algorithms, etc?.

G

So some of them just don't really work for our problems. We have scalar and non-tailored values that we're picking, um I don't know their reach. You know why.

C

So we've initially started with something like simple grid, search or variance of that, and then I think we had a summer intern who said.

K

Hey this is really done. We should be doing something.

C

K

Than this, and so he.

C

K

Did some research.

F

And came up with.

C

A particle swarm optimization as a really good fit. He implemented it prototyped it and showed that it created good results, and we said great, so that's sort of how it happened, but there are actually some good reasons for choosing pso over some of the other techniques, not.

G

To say that some there won't be some other technique that.

C

Is even better, but the particle swarm optimization is a very nice combination of doing local search and global search um and in our.

F

Situation that seems.

C

To be a really good fit because in some.

F

C

The optimization.

F

Surface is very.

C

F

C

F

Places it's really flat.

C

F

Want to be in the flat areas as.

C

Possible but pso is a very nice. They did a good job of combining local and global search where something like um gradient descent. Only.

F

Does a local search.

C

And um yeah, so this is a nice combination and I think simulated.

F

Link just does a global search.

C

Doesn't look at the local shape of the surface? um It's also in practice. uh It converges pretty fast, um so simulated and annealing again is is computationally a lot slower than pso and it's.

F

Also, very paralyzable, so you have lots of particles and.

C

Each particle can be computed completely in parallel and then you combine stuff and then you, uh you know, run another generation. So there were a bunch of good properties that it had as well, and I think.

J

Scott also mentioned.

C

J

K

C

A categorical data, so we needed a technique that could handle both as well, whereas something like grading design might not handle that very well.

D

And I think uh the second question here, which is.

A

D

We experiment with other techniques I mentioned. We started with a simple thing on our own uh and I believe some other people did some other somebody else did uh um early on. Did some other tests remember that.

F

D

On um and I I recall, some other people doing some quick experiments using uh genetic algorithms viruses and um I thought my recollection.

C

D

They find results, we haven't.

C

Spent a lot of time trying.

D

uh Trying to really optimize that process uh we've got good enough results. I think we kind of stopped there a bit, and I guess the third.

D

Question yeah: the question which is you know: are there some generic model parameters and I don't? I can tell the story or something else, to tell the story there, but they're wrong. Okay. Well, so he was trying to follow this. um So you.

A

C

D

And you're running on some number of parameters- and you end up with some answer for some particular data system for this data set. We would say this that parameters leads to the best prediction result, which generally is also the best sonoma detection result and, and then we'd run out on a different side of the data. You get a different set of parameters.

D

So then, again one of our interns, I believe, an intern, a different intern. Yes, did a little clever thing.

H

D

uh Well, maybe we asked them to, I don't remember, but um we looked at. We took a whole bunch of data sets and we and then we looked at the sets of parameters that each the swarm came up for each one of those data sets and it turned out that those the the ideal parameters had a clustering to them. There were about, I don't know five or six clusters or something like that, but you chose five thinking. Oh he could have caught. Maybe he chose five, let's say so, but there were there.

A

Was definitely a clustering.

F

D

They tend to be in different clusters and so that's cool, so maybe we don't really need to do a full swarm. We just need to figure out which one of those clusters was the best cluster.

D

Then we discovered something else which is really really surprising, that if we took the centroid of those clusters and use that instead of the ideal parameters that the swarm came up with, so this one comes up with the best set of parameters it can find. We say no.

F

Don't do that, let's just say the centroid of these different.

D

Clusters, well those sense rights work better than the actual output of the swarm. So this is like wow, that's great, that's surprising! um So then we basically have something to pick 105.

D

and then then we found further that when we were developing gronk, uh at least for the survey date and all that we had a bunch of other data, sets too it wasn't just the server date. But a bunch of data sets that we found out that one centroid of one of those clusters seemed to work well on the vast majority of the problems we're looking at.

D

Not all of them there's a few that it didn't do as well on, but it was a small percentage, and so we in the end decided not to do any forming.

D

We just said just pick that centroid for that cluster and use it for all of our problems and that's certainly a lot simpler and it made everything a lot faster um and we haven't.

E

D

E

And where is that point.

E

C

Those parameters are checked into.

E

C

Example- um and I think, there's been email trends around this topic as well. If you are.

F

Working on a data.

C

Set that is very similar to the energy.

F

Data set that we have.

C

The hudson data set or ip.

F

Data sets they all have this sort of time stamp.

J

C

You know scale of values.

J

C

You know not too bizarre, then you can just use those parameters and be fine with it. If you, if you work on a.

F

Problem, that's very different or has lots of different other.

J

C

Coming in then, you might want to use swarming to come up with a good set up, but then once.

K

You have that set.

C

You probably don't need to keep restoring.

D

K

D

Geospatial tracker: did we swarm right? Let's use the knights in so there.

J

We go so the whole everything we do.

D

Is another example: we did the whole geospatial tracking thing with.

J

Gps encoder and we didn't even bother to swarm. I loved those caveats. I did change the learning rate, changing the ratio of increment to decrement for the tempo memory, so you did a manual last year.

D

I did a slightly.

J

We did that because we got better results.

D

So you're, just so patient.

D

C

So I think, a lot of in a lot of cases, the hot chin data, the parameters that are in their work is a very good starting point.

D

Yeah, so it wouldn't.

C

Be the first thing I would worry about just like.

D

We did the geospatial tracking, we didn't worry about that. We had a lot of other things to worry about. We got decent results with that and I suppose, if you really want to go commercial and if you want to do.

D

But you know I don't know if it'd be better or not, but I wouldn't be surprised if you said.

C

I did a lot of work and.

D

Got marginally better.

C

Then there are other cases like.

D

The audio stuff.

E

That matt has been.

C

Looking into there, you really want to swarm because you have potentially dozens of fields. You don't even know what combination well.

D

That's that's a great distinction. If you're trying to decide which fields are contributing, um then you really do need this one. But if you're just saying I'm just going to build a model of this, this data point and maybe with time, then you can start with the preset parameters.

G

um So I also put a couple links in the wiki page and the question that are just um a couple papers where they evaluated different, optimization algorithms and in both those cases, uh pso came up as the best one, and these.

D

Are actually just ones.

G

The only ones that I found that compared different ones, that included yes for other.

I

D

Just saying in general, psas are pretty good.

G

Yeah yeah. Yes, it's a really good story. I think that the main thing right here is that we're pretty confident that our prediction is pretty good. You know yeah, you might be able to do better on some problems, but it's good enough for you.

C

G

Is a very good start: pso is a very good starting point.

C

Usually much better than simulated and legally more genetic algorithms.

A

And if you're listening and you don't know what pso is ron, one of our former engineers explains it in the swarming.

F

A

Youtube our youtube channel, so you know: okay, let's go to the next question, which is back up here: pattern stability. So I don't know how how well we're going to do to answer this question. But there's been this really long discussion on nuke theory, uh mostly between uh virgo and rob freeman, um face face dong. I don't know how to say that and uh david ray about pattern. Stability.

A

um Honestly, I haven't followed the whole conversation, but um I asked david to try and give a little summary of the questions he has based on that and he broke out two questions. One is given prior training for the sequence of of uh characters, a b c d, e, f g h when starting a fresh input at d. How does the upper region stabilize on the abc abcgfgh, when all they will see, is uh d-e-f-g-h well.

D

This question I mean again recap right here: I don't think any of us are really understanding the depth of this threat. There was a lot a lot of textbooks.

D

Honestly, we didn't all father so, uh but if I.

C

Understand this particular problem. I.

D

Could take a stab at maybe someone else has a different interpretation problem. uh If you have a long sequence, it sounds like they're asking well what, if you start a middle sequence right, and uh this is how does it stabilize it and what stabilize.

F

D

That it's it's they're asking: how does the temple pool perform stable representation? Well, the timber pool or former stable representation if the sequence is being tracked? So really? The question is: how does it pick up that sequence in the middle, and that was one of the very very first things we tried. We made sure that the the temple memory algorithm did is that it needs to be able to pick up.

D

You know what melodies, if you want to call that in the middle of the melody you know they pick up sequences in the middle and so at.

E

D

In time, if you have an input in this particular case d, um the d is out of context you're starting fresh there's, no context for it, and so basically, all the cells in the columns that represent the equal first and essentially, at that point it tries to predict everything.

E

D

Has d anywhere in it and it'll say: okay, what you know d sometimes followed by e and sometimes followed by q, and sometimes followed by z12 or whatever uh and then and so, but then on the next input. It will very quickly settle out it's all like named mattoon as soon as it starts each time that the predictions get narrower and narrower and very soon says. I know I'm in at that point once you've once you've zoomed in on, like it's going to be this sequence or these two sequences.

D

If possible, um then the temple pool should start working and it'll become a stable pattern. So that was there from the.

F

Very very beginning.

D

And um it's just part of the way the temple memory works. Did anyone else have a different interpretation for them? I'd just say bursting it's what we call a minute yeah, the first things about all the cells in the column. That's basically saying I'm getting a d, but I don't know I don't have a context for it. Therefore, I can't put it accurately what's going to happen next, but I operate everything that my it's very very.

J

D

The little of the game name that tune reversal. You hear the first interval.

E

And you're like well.

D

I have no idea- maybe I do very, very second and by the time you get third fourth interval.

A

Okay, this next question is given two patterns: a b c and a b c d e f, starting on abc. How do you get it to discern whether it needs to keep matching and extend the recognition's abcdf when it may not be desired? The second was product by.

H

Jake first round they've.

A

H

About it, through this level, threat.

A

D

D

um When do you assume that you're trying to learn to extend the sequence, and when do you say uh no, I'm at the end, and I need to recognize something else. This is all the.

A

Whole, um I don't know, I have to wait for david to hear that, and maybe he can clarify that this is the. What do we call those terms.

D

David's watching.

A

So you know um yeah.

D

Pam I mean this is this: is I see if it is the problem or thinking it is? This is a very serious question. It's it's kind of difficult to say to tell the memory like you know what you're done started looking for something new or no keep fix, try to try to learn this as an extension of the of the existing sequence and it's difficult to do those both at the same time.

D

That's the problem which I don't know if that is a promise, we spent a.

F

Lot of time on.

D

This, in the end, we've kind of concluded that we we did it. I think, in the code today.

F

We have sort of.

D

I would call it a hack, but it's kind of a non-biological solution to this problem, where we we try to extend sequences up to a certain point, and then we actually also backtrack and try to pick up patterns and see if something else could match the sequence and which I don't think is biologically correct. But it works pretty well, um but in the end in biology I think what's going on, is your brain.

G

Is it's almost like you're, either trying to learn or.

C

D

Trying to infer at this point in time it's kind of like you know what what's your mode right now, if I were to tell you um uh imagine we'd, go back to knowledge as an example, and I said, okay, your.

F

Job right now is to learn new.

D

Melodies, uh you will be in this mode of trying to remember these actual patterns. If I said to.

F

You your job right now.

D

Is is to just you're going to get you know, life or death situation. You have to recognize the knowledge you already know. You won't be extending it. So there's sort of two different ways: you're dealing with this situation, one is you're trying to extend and learn, and the other is you're trying to just confirm- and I think that's a little bit more. What's going on in biology, it's a little bit more mobile, but right at the moment, we don't do that.

D

We do a non-biological trick to try to do both at the same time which works. Okay,.

A

And that trick is currently in newpick. It's not it's yeah.

J

I don't know if it's in the new jump before the it's, not in the new implementation of temple memory, yeah it's in the current, it's in the one that, uh like tp.5.

C

Yeah, so in that case it would just take multiple repetitions.

D

C

It's a little bit more involved.

D

When it's over here, but it can, if you try and extend.

I

D

It's similar to other sequences. It can take a long time to extend it uh to learn it. So there were two things: it was pound length and there was backtracking. I think that term was.

D

And then backtracking is like okay.

F

Now I've tried to extend.

D

My sequences, let me go back and see if something matches you know, try to catch up.

G

My answer this would have been you're, always trying to learn the new things that you see and if you see things, if things happen, multiple times then you'll hit whatever the threshold is for it to be connected. So.

C

I don't know it. Doesn't yes but yeah. That's what would happen without the pan mode is every time you.

I

F

You're always trying to extend it a little.

C

Bit but it's not it's not strongly.

H

C

Each time so you would have to see it a couple of times and then.

F

You would extend it by one step.

C

And then you have to see to see the whole thing it takes.

D

A long time it takes a.

C

Long time, but that that understanding is correct, we.

D

Were trying to speed it up, um so that was it, it was just. It was taking a long.

J

Time it's not as obvious in the abc and abcdef case, as was the question it's more obvious when you have a shared sub sequence in the middle of a sequence, so you have a sequence that you've seen and it's a higher. The whole thing is, however, sequence, but there's this the subsequence in the middle is shared among other hierarchy sequences.

J

Then, uh if, if one, however sequence is blurred already, then a different hierarchy, sequence that contains the same sub sequence, you'll need to pass over it many times in order to learn it as a new height of recipients. Otherwise it keeps falling back into it'll make already it's.

D

Going to constantly try to say.

I

Oh I've seen this before I've seen something like.

J

I

And then I'll several passes for each for.

J

I

Of the pieces in the subsequent yeah each pass, it extends.

J

By one, so it needs to extend so it needs to make those many passes to extend by one until the end of the.

C

There's some biological evidence for this, something.

J

Like the attention.

C

One where you're focusing on something it doesn't take so many passes. It just extends.

I

It very quickly.

C

But if you don't, then.

I

It takes so long, I'm sure.

D

I'm sure it's the case again going back to the uh to the melody.

I

D

It's sort of like.

I

Saying, okay, you learn all these melodies that have.

D

Common self pieces to them, and now you're you're, going along and you're following a melody and and um you're going to constantly try to match it to one of the previous ones. But even if I start in the middle of the melody, oh yeah, that's that's why 21.

D

and it's just going to always track things you know, and it won't it won't say I'm experiencing. This is new I'll, try to learn to do higher receivers, so um you can sort of see all these things and the analogies that help. You really see these things, but I think.

F

I really do think in biology. It's.

D

From these two different modes, you can say all right: I'm trying to learn something new it'll run quicker uh and there are. But there are chemicals that release in the brain. That say, you know you're trying to learn so, okay.

D

um That was really what all those emails.

A

D

And that's great, but they should seem more complicated that when they're online.

A

They were, um they were heated at times, yeah, it's a difficult problem, yeah.

D

I just didn't even understand.

A

I invited fergal and rob to this hangout just in case they both wanted to get in the same room and work it out. Maybe we could do that as a another event and charge for tickets.

K

That's what they tried to do last night.

G

Do you think that some of those things like um like pam and these different ways to make it learn faster? um Do this make sense from an information through your perspective, or is that only when you have some knowledge about the data it gives me something about data for those to make sense to use? I don't think so.

D

Yeah I mean I mean they apply to all data sets. It's not like. We said. Oh, these are only working. Some data.

J

But is your question that is that signal whether you need to learn more or you need to infer infer more needs to come from outside of the dataset.

G

That's one of the parameters that actually changes and can have a big impact on the results. um So.

C

Yeah, I understand I mean it's hard to say.

K

C

K

Setting on pam, I can probably come up with a set of.

C

Sequences which would work against that setting? Yes,.

G

C

So you could, um I don't know if there's a priority, if there's some perfect setting, it would be well.

K

D

C

More of a practical question, given realistic data sets, what do you I.

D

Don't think it comes from the data set itself, but that was like, I think, in a biological point of view, it would come from the data it would come from. Some external motivation, like your job, is to.

F

You're going to get a reward.

D

From memorizing this list of patterns versus you're going to get a reward for uh recognizing these patterns, it's a very common psychological test. They do, though, I get the name for it, but they, you know.

F

You're, given a list of words, for example, you have.

D

To memorize them and then later you have to recall them. Well that your mind is in a different set of framework when you're told your job is to memorize, and then your job is to call them and yeah and there's all these.

C

Attention blindness stuff yeah if they give you.

D

C

Stuff, you don't remember that.

D

This is a complex part of brain theory that we don't really.

F

Deal with here.

D

Which is that uh these are the uh some of these? These are chemicals that are dispersed broadly throughout the near cortex, that impact learning rates and attention and various things like that. We don't um we don't. We.

A

D

With them at all,.

A

So pam stands for pay attention mode and is that a term that we've manufactured ourselves or is this.

H

A

H

Probably some imagery.

D

We're suggesting it's an intentional mechanism.

A

Study learn this. I need to find that switch on my children.

H

A

uh Okay, um it looks like we're out of questions. uh We've got 50 minutes or.

F

So for this office hour, so that's great. um It was really nice to have some of these questions ahead of time.

A

Yeah I'll do that next time too. That is helpful. um So thanks everyone for watching, thanks to our panel of uh new big engineers, for answering our questions and we'll we will be back with another office hour. The when I say second monday and then.

A

Thanks for joining in thanks, everyone.