DevoWorm ML, 2 Oct 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DevoWorm ML: Week 5 (Pre-trained Models discussion)

Description

Fifth DevoWormML meeting, October 2. Attendees: Bradly Alicea, Jesse Parent, Vinay Varma, and Abraham Kohrman

A

All right pretty good! How about yourself, oh good, I've, been going to all.

B

Their papers that.

A

You put in a sink that.

B

A

Some other stuff.

B

So I found some interesting things.

C

Oh yeah he's gonna be.

B

So I'm kind of exploring that, but more about that later.

D

C

Good yeah, I guess I can start people may be coming in later, but uh first of all, uh how are you guys doing so far? If you find anything you want to follow up on that? We've talked about in the meetings.

D

C

What I'm not familiar with what your specialty is, Abraham you. What do you want to do like with like machine learning specifically, but what's your research boat, yeah.

C

D

Analysis so I be on across tepee and I.

D

E

Sanitation problems, the.

D

E

Problem that we think is probably for.

D

The intractable.

E

B

Cracking itself and.

D

Then that provides us.

B

D

To behavior over time, and then we have a a time: alignment of fundamental data problem, pull up by some spaz and it's a very drying.

B

Here sigh anyway, I'm talking about exactly.

D

What I'm doing, but the actual, exciting biology didn't a lot of us is we're able to see all the reasons cycle in living animals and correlate it to someone else behaviour, including the differentiation of complex all types. Oh that's.

C

Good yeah that sounds pretty good, yeah, yeah I. Think cell cycles really neat area like just in terms of like understanding how cells divide and you know the timing of it, especially yeah I used to work in the area of so cellular, reprogramming and I was like one of the main things. If you want to understand how it like how cells sort of you know reprogram, one of the things you have to understand is like how they divide and like what happens. You know when they decide to.

C

You know, convert to a different fate or not yeah so and then so you're just kind of working on the cutting edge of that area like they take it yeah. So that sounds pretty good. Maybe at some point you can give a short you know, presentation and if any, even if it's just on what you do, that would be good yeah.

C

Let's see, jesse has a comment mostly feeling my word way towards a specific ml project when that said on one yet, okay, so I guess yeah. Let me talk about the pre-training models, so I mentioned in the email that I was thinking more about pre train models and, as we talked about last week, we have these models that are, you know, kind of we. We want to use them to sort of classify training data so that we can build a nice model for some problem that we're interested in analyzing data.

C

That's you know more diverse than the training set, so that I think it's worth like going through the idea of pre-training featuring models just to kind of get an idea of. What's out there.

C

D

D

The value the training time that it takes for the network to learn what an image is going to recognize new types of shapes to thicken types of shapes and to be accurate at that task was much much faster link. A network.

E

D

A concept of images contains structural and shape information, so retraining from a network that has already developed the ability to.

D

Perceive information out of images is a month off thunking tasks in terms of computational time, then starting from scratch.

C

Yeah, actually that was might we're gonna mean my point on this. You know I think that one of the problems we have, of course we saw in the digital vessel area stuff, so we don't have good models for biology. We have good models for like will same language and, like you know, maybe cars or you know trees, but we don't. You know, maybe more general problems too. Well, we don't have a good set of models for biology, and so it does require I think specific sort of model.

C

You know sort of a set of assumptions that you wouldn't find in a more general model.

C

So let me let me go to the slides, oh sure, my screen and you'll find this interesting to Jesse. I think talk you you mentioned you're interested in a little bit when we talked on slacks.

C

Let's see we can, let's just go through these and have a discussion about that. Can you see my slides, okay, good, okay, so.

C

This is like pre trained models for developmental biology. That's what I'm calling this and in this image here is actually a sort of a graph of a lot of the pre trained models that exist and they're in this. In this case, they're measuring the number of operations it takes to sort of train one of these versus the total accuracy.

C

So you know with these pre trained models, you can throw them at a problem and, like abraham said you know you can it might take a while to train it, a good accuracy, but you also have to worry about it over a fitting or under fitting your data. So you know if it over fits the data, it will predict. You know, you know predict that anything is a feature if it's under trained it won't find any features at all in your data.

C

So you know if we go back to the the spherical systems that we talked about last week. So these are two embryos.

C

You know it's C elegans embryo Genesis, and this is your soft lumber Genesis and you can see like the the embryo genetic process, is very different in these two organisms in Drosophila you get this cellular ization and you have many more cells that are, you know, forming some sort of shape. That's going to become the larval stage, C elegans, you have the same process, but it's largely cells dividing and moving around, and then they form you know, wait. They have like a comma stage and they have other stages where they take.

C

You know their shape is defined later on, but all of that I mean that requires us to characterize movement. That requires us to characterize like one feature becoming two features and also it requires. You know some standardization of the process.

C

So if you go from embryo to embryo, you know it isn't it in C elegans, for example, it's highly programmed, but it's still, there are still variation in the process, and so the variation is a typical say, like a bunch of phases where we can characterize human faces may get a pretty good handle on you know a face is a face, but with an embryo we don't really know that so I mean you know. We talked about pre train models and you know how they might be used.

C

It's basically taking a say, neural network and defining the parameters in advance like some of the connections and and then people can use them to train their. You know, put training data in, and you know kind of get an answer for their problem, but in biology it might be will be tricky, but one of the things that I was thinking about over the week is that we have models of biological processes, especially models, an app like a I'll call it a neural network. But one of those is the perception of biological motion.

C

So this, in this case, in biology, we have motion in the embryos or you know, whatever we're looking at, we usually do have motion in biological systems because we're looking at processes for this. In this case, we actually have a model of perception perception of motion and, and what this is is it comes from psychology actually in in the middle of the last century, a number of ecological psychologists started looking at biological motion. So what this is is well. If you look at wave, look they've been walking individual like in this case.

C

You have two people walking across see on the right. We don't recognize like the whole of the motion. Well, we recognize you know we're recognizing the whole of the motion, but we're doing it in a way that, like focuses on specific parts of the law of the Walker and so they've boiled it down to like these points.

C

So they're just like points that you might see like the Bini, the foot, the hand and they tested this on people and people were able to recognize that as two people walking, even though there's not really two people walking, it's just two sets of points that are mimicking the walking motion, and so people have done this sort of experimentation for a long time and they've gotten to the point where they've understood kind of a model for what's going on in the brain.

C

When we do that and so they use, then this is very similar to what people do in machine learning where they pick out features of the object you don't necessarily have to analyze the entire object. You just pick out features and you relate them to one another in a way. That's like makes sense of it. So in this case we're looking at a cognitive model and in a cognitive model, you have some, you know, input some sort of sensory input in this case.

C

It's the vision of like this biological motion and you pass it into some neurons, which are which extract the different components of motion. So you have maximum and minimum sets of features, and then you integrate them, and then you represent these in a certain type of cell, that is called a template cell, and this leads to two separate decisions derived from a compact representation.

C

So what the brain does is it represents this information in this way that we see down here these different points that are moving around and then that information gets integrated into an entire. We know you know we recognize human movement because we recognize those cues and their movement, and we recognize that as as movement, so you could actually create a fake situation where it looks sort of like a human and you might still interpret it as a human moving.

C

But it's not a human at some other thing it could be like maybe just a set of wind chimes or something and people actually you know, but, as you know, people miss classify objects, maybe is human or animal motion. All the time, so I mean that's one.

C

One area that I kind of get I just thought about when I was thinking about this problem is that cognitive models for this sort of thing might be useful, and but another thing we have to think about is that there is this idea of process models where people have looked at say, like pre, train models, they've looked at neural simulations and they've, looked at cognitive models and they've thought about like how you know how they map to cognition and how they map to biology.

C

So a pre train model, or even like a machine learning model, is mimicking the brain in some way and so we're mimicking the brain in cognitive models. We're mimicking the cognition we're not necessarily mimicking the brain. So if you look back at this model, we have neurons in it, but we're really interested in the representation.

C

Conversely, something like the Blue, Brain, Project or hodgkin and huxley is where you're modeling a biological neuron, but you don't really care so much about what it does in terms of cognition. But what's interesting here is that these pre trained models and deep learning models sit kind of in-between this, and so they sit in between cognitive models and sort of like you know, high fidelity, biological models, and so you know for something like biology. It might be interesting if we sort of you know adopted.

C

Maybe some of these attributes as well, where we actually maybe used something that resemble more of a cognitive model than a typical pre tree model. I only say that, because you know we know how to interpret biology. We in developmental biology, there's a history, people kind of looking in under microscope, slides, and you know interpreting things through drawings or through some other hive Annette. How are you.

E

C

I was just talking about three train models and I was entertaining some thoughts here. Talking about.

E

F

Right now, I'm talking.

C

About something.

F

C

F

C

Is taking the keep, learning and pretreat.

C

A

The models of neurons.

C

And so I was mentioning that we may be able to do something interesting by mimicking cognitive models rather than something like a typical pre-training like I'll make these slides available to people. We can have a discussion about this. You know we could do it here and then we can, this later, so I didn't want to get people's feedback, so this is sort of paper called cognitive, computational neuroscience. So this isn't from like the machine learning literature. This is more of a cognitive neuroscience approach, but you know that there are ties.

C

I wanted people to be aware of the ties between the two areas.

C

So why don't we go through some some pre training models that exist?

C

There is one that just came out and this is actually as it's very closely allied with biology and that it's a deep learning kit for looking at animal poses and movement, and so this is deep pit, and this is something that there are people in here who have done like collective behavior and emergence type. Work like e'en cruising does a lot of stuff with like bird flocks, so I mean that's.

C

Where I think these people are coming from on this uh they're they're from like an animal behavior kind of background and they've developed, this kit, which is a set of animal, poses and they're able to estimate movement from that, and so this gif isn't running here, but it shows a bunch of things moving around and they're able to predict the movement from that, and so you know in in this case you're not using a traditional pre-trained model but you're using something that's inspired by sort of what we know about animal behavior, and so in this case there's something called a stack, that's net bottom and so they're, using this sort of process where they're taking input, images and they're augmenting them using augmented data.

C

So here in the Augmented data augmentation strategy, they have this input of a fly and then they have it in different poses you know, and so when the fly is moving they're in these different poses, you take static images and put them together, and then you run this Network there's some math here. The basic idea is that you're able to take, you know, features and a you know.

C

Some of these poses and put them in a graph and then get some sort of prediction of movement, but the movement you know is based on sort of what we know about the biology, not necessarily just well, here's a model that you know works on language processing or on you know, classifying cars and we'll apply it to movement. So they took the bull of a different approach here, but still this isn't a cognitive model.

C

This is more of just a regular sort of you know, machine learning type model, but these they sort of informed it with biology.

C

But then there are also other types of free training models. One that's has a very lengthy blog post behind it is I, have a GP to mock EPT to model, and so this is an example of a pre train model for language processing, where they've been able to train it on language processing. But you know if you use it for another purpose, you know you might not get the kind of result that you want and again it's like very good for its test domain, but it's not enough, doesn't necessarily transfer very.

C

But if you want to understand pre-training models, that's a good place to go to to this blog post they'll. Explain to you some aspects of it. You know how it works, how they train it with data and so forth, and so in.

D

This case, you find that a lot of things.

C

You know they train it with a lot of data. First of all, so you know it takes a lot of data, but there might be cases where you're creeped retrain models that don't require a lot of data. In fact, you know it depends if you're looking at a problem biology, that's hard to characterize.

C

You know you might not want to have a lot of input data or rely on that, because uh you know it's you just don't have the data that you can throw at it and this and this blog post. You get an idea of the scale that you need and it's it's a little intimidation, but that's that's how they that's what sometimes they need to do to to make these pre train models work.

C

Another example from language processing- and this is an area called NLP, which is your linguistic programming. It's called fast bird and this is a simple, deep learning library. So in this case they do two steps. They do like a semi, supervised training step on a large amount of text, so they just you know, throw a lot of data in it and then the next step. They supervised the training, they label a data set. They put that into the algorithm and they're able to benefit from the semi-supervised stuff. So semi-supervised means that you have text.

C

That's you know. Sorta has labels and sort of doesn't you're, throwing you know words at the bottle. Sometimes it knows the context. Sometimes it knows you know. Sometimes it knows the meaning. Sometimes it doesn't they're allotted you know it's it's very diverse in that way, and then, with the supervised training, you have something with a formal label and then from that they can create a tree train model that can encounter a lot of different cases. But in that you know in in this case again, you need a lot of input data to do it.

C

So that's all I have for this. I just wanted to go through a couple of pre tree models and get a little bit of feedback from people as to what they you know. Maybe they they think about like pre train models, maybe Jesse actually mentioned he was interested in. You know pursuing that a little bit further, at least in terms of thinking about, like you know how we might create one for biology. So.

D

And I think I think there there are bunch of different tasks that would ideally want to be able to do and different types of network architectures are going to be very different types of tasks right.

C

D

So you have, on the one hand you have human ossification, which requires internet education right as well and in in general you will also have simple sanitation problems and I guess there are a bunch of other types of ruse that you could. You can attack with no networks pretty effectively and I think there there are.

D

Generally, there's a lot of.

D

For these different types of tasks and.

D

Yeah, it's made a big problem, often I think becomes which.

E

D

Articles do I try for this tap and you know that the data hungriness is a real problem. Clearly, when you need to handle your data and that can take really time which is maybe having a a future.

E

In architecture that motion images or knows what.

D

My data type is, can be an absolute lifesaver but having having email that simply is able to understand the structure of the information site that you're eating a already increases the amount of day defeated by like an order of magnitude, and that makes things a lot less painful.

D

It's a little definitely value in coming having free training models, letting is also a question of what is the. What is the precise test, but you want to tackle.

D

Do you want to attack? All motion will attack all feature, detection, we're gonna tackle identification and the horizontal segmentation and the architectures that are most.

A

Effective at each of those four tasks.

D

Are currently different, we don't have a general-purpose nor network structure. Yes, that'll do any tax key, throw it in the thinkin I, don't listen! You know that my work that way. Either we use different network structures, different paths and then we supervise which one we're passing it to us. That's sort of our brain works. It's centralized, but it's still solely attributed processing. You have these regions on the cortex that are associated with different behaves right. You have a sensory monkey. That's this way right on your brain cortex and you have a visual process.

D

Okay, all these other structures that are associated with specific types of processing, tasks that have the appropriate network architecture for those particular tasks and then there's some sort of classification that passes the problem to the right, neurons right and so at the end of that I. Don't necessarily know that you can get to a general purpose solution by building a single Network, so much as building a bunch of very specialized networks and figuring out a way to figure out when you want right now. The problem is I.

D

Think the very entry for a lot of us. Other than me, the density of the descriptions and the.

D

Options that are out there is understanding which model is most likely to behave well on our data. That's that's a real challenge because we're that classifying function that decides what our data type is and what malls, when you look the best yeah.

C

Yeah I think you're I think that's right, I think there. Definitely it's and you know the whole idea of having a general model. I. Think, though, that's what's assumed, you know a lot of the pre train models.

A

C

Know we kind of increase one model and there's a there's: a universal.

E

C

There you'll be able to turn I mean that's kind of the principle behind statistics. You know that you can take a measurement and then apply statistical tests and then no matter what your data is. If it's like you know, defects in manufactured goods or like people in a sporting event or, like you know, cells.

C

You know it's just like a broad range of things that you could measure, but you should be able to use this one set of tests. I mean now. You know there are a lot of statistical tests, which maybe suggest that you know that kinda universality isn't really doesn't really exist. There are a lot of assumptions that need to make, for you know different problems.

C

So yeah I agree that you have to. You know that I don't think that there is necessarily a general solution, but that barrier to entry mentioned is actually pretty important to I. Think you know, I've seen a lot of people, they kind of jump into the myself included kind of jump into the area, and it's like you know what is the best model for the kind of data were using. That's why I did this I wanted to focus on this a little bit?

C

Because you know we don't really know what the best model is, and we kind of just have to use trial and error to get there.

D

Yeah I think that you know on the general.

E

Society, linguistic, a you know: the unit will not be able to do anything.

D

Because the dependent all to handle- because they do have a strong future and is multi-dimensional array- is.

E

D

The Uni that.

E

Does a really good job of segmenting images into.

D

Program processes in the fire sent to.

C

D

C

We have to be armed well.

D

I'm going to get a nothing out because modeling is a really good job, that processing language data and yeah a lot of time with Lee West, because.

D

D

You know that will pay. My view are very different.

C

Sorts of tasks.

D

Like a questions.

C

E

C

Apology I'm, like okay,.

A

Well, we're in biology Thursday, but.

D

E

A

D

To ballistic data and the closest that I could get for that was well, we got a sequence data which is the Raymond urbanised, is one the benefit and you can in principle break it up into units that are working in our genes and promote a very alarm, contributed language, that's us, and then you can attempt to use that power to your first problem of person of information. But it doesn't it's a very.

D

C

D

C

I think that's I think that's very, very good observation to like you know. I know this that, like in very early bioinformatics, they used to focus on you, know, gene sequences and then, like you know, you can use a linguistic models like a lot of people like to use hmmm models which are hidden, Markov models.

C

You know that are kind of for both you know linguistics and, for you know, sequence, data and it works. But then you know it depends on what you want to get out of the sequence data. If it's you know it's, it's not really structured exactly like language, so you have to find a model, that's sort of abstract enough to deal with that. But then you know there are certain assumptions of the model that you know: don't necessarily fit the data, so it's always kind of like you're.

C

You know in and I think a lot of times in bioinformatics. In particular, they try to modify the models to sort of fit into the assumptions you know so. There's, like you know, they're like platforms like you know, they might use a hidden, Markov model, but they might have you know they might tweak it to conform to certain assumptions about gene sequence, data that sort of thing. So it's it's.

C

You know it's it's good to kind of look at what other areas of modeling you're doing. Actually you know you also have platforms like go back to agent-based models where, like you know, you get a package that does agent-based modeling often times they'll come with templates which are like you know they have different sort of scenarios built into them. So, like you know, you have an agent-based modeling program and it has like models for, like famous, you know, cellular automata models or they might have models for you know different game theory, game game. Theoretic games.

C

I was just working with a package we're looking at like like prisoner's dilemma models, so you know, you'd have a agent-based model and you know you go in and you'd say well. This is a good. You know you know so I guess a sub model for what I want to do. I want to look at this problem in terms of the prisoner's dilemma, and then you know you might have variations on that and it's all so pre-code.

A

C

So it gives you like a model to work for him but thinning. You have to make sure that your data has its the assumptions of that monitor and that's the hard part is getting that right because you know sometimes, if you violate this I'm sure you get a result exactly.

D

This is sort of a joke.

E

But not entirely.

D

E

An entire academic field that does this.

D

All the time and it's economics- oh yeah, I'm.

C

Gonna go take these.

D

C

Flow equations.

D

And using the model human behavior by removing half of those variables- and you know, there's precedent for doing this sort of a test, but at the end of the day, as we're hugely shown, those models that you know a lot of patients don't to attract leakage on a model include behavior. You know, because there's all there's a lot.

B

Of repurposed hardware type problems.

C

D

C

You, when you try to make all the best economic models.

D

You know are based on you see a flaw balls right now and they're terrible yeah. What the modeling what's been happened.

D

C

Exactly so, yeah I mean that's like a like a whole area of like inquiry. I guess I've seen a couple papers actually come up to talk about. Like many people are kind of questioning your assumptions where you know you find something. That's universal, like I. Think power laws is one example from like complexity theory. So about 20 years ago. Everyone thought that there were power laws and like networks and in you know all sorts of things and then, like you know now, people are evacuated insane.

C

Was it really a power law or is there more to it, because the idea of a power law is that you have this one statistical signature. It's supposed to be. You know, live alone across all these different domains, and then you know is it? Is that really the case or are we just kind of looking at many different things that are very similar, but we kind of use a statistical model to you know claim that there are these deep.

C

You know deep fundamental things that all these systems share. It's you know it, but it is relevant to like applying. You know your data to a model because you do is make you do have to make that assumption at some level.

C

Like I said, you know that, like there may be cases, there may be opportunities to look at like cognitive models, which are you know, models of how the brain, actually, you know, processes like, say, bio, biological information as sort of a sort of a key to like, maybe how that can be done. You know in a more specific way.

C

You know the problem is, is like you know: how do we? How would we code that and how would we like throw data in it? Would that be something that would be feasible? I know people are doing like using cognitive models. Sort of to look at you know you know maybe for like data analysis, but that's you know it's so models even like that or a little bit tricky to implement.

C

So I guess just even a if you have any thoughts on this.

E

E

E

The same experiences sometimes work very well. Some parents may be my choice of majority. Selecting of the model was bad, but I see how experiment a lot of things on there, a pjr for picking on us and so I think this. This useful people models. We encourage because a person looks like Elsa flow and pipe ouch had these humps.

E

If you were bitten, they have.

F

A set up for all the people as on.

E

F

Tasks, first of all, sedimentation for the exchange in and so on. So.

E

I think it's a.

F

Rising used to practice and.

E

If we have the right amount of data and I think we can use one of these models and it's one of these models and see how.

C

The results are there yeah, so I see that the tracks are fall as a new version out.

E

2.0 alpha and estimated release like the.

F

Stable version got.

C

E

So they have a lot of missing in the previous mix of conventions. Then the nurse will be a case. There's a TF naught M L and D F da cleared sandy hook onto the dark layers, so there are different. It is so. What the major update to what I see is that they remove all of those things and the bonusing will be a recipient, and also that is remember it so.

A

E

We need to expand principal and now.

E

Actually, so that that is the major update and all the documentation has been done so I'm really excited to.

C

E

C

So do you have experience, of course, in like dialog like in the working with the embryos like if you worked with what kinds of different data, if you worked with, if you use pre train models for.

E

C

Yeah I think we can trust them yeah, it's just a matter of like you know. If we have a certain type of data, what are we using something that maybe is trained in a totally different domain? And then you know I guess the first question is you know? Maybe is there a domain of data so like we have like? We got, isn't universal necessarily there might be some aspects of it that you know that explain.

C

You know there are data types that have very distinct things. So, like you know, if you had a bunch of M&Ms on a table, we took the picture and you asked the algorithm to pick up the M&Ms. You know if you used like, maybe like something that was specialized from linguistics. That may not be very good if he picked something. That's optimized for object, detection.

E

C

Something color changes, then it would probably do a very good job of picking and so but then you know, that's like a very general level, but still you see the difference so I think that's actually important to his understanding way.

C

C

E

In those object, so what we did was we remove the last? Yes that please, like it has some 50 objects, I think I, don't think I, don't remember the number. Also we remove the last way my fees of the model we remove the last years and we are didn't move with a new class. That is a cheetah that good yes, so we found this on. Jobs like this is a very I said.

E

Custom data so and then I tried it so when I found it to be as a model because we trained it for a very long time and we had very nice data, so I think with the custom later yeah.

C

Yeah I think that's yeah. A lot of those kind of strategies are gonna, be useful, yeah. It's very interesting. um Let's see, Abraham I think posted a link to paper William Bialik who's, a famous physicist dimensionality and dynamics in the behavior of C elegans I. You pick something that is even worm, approved, I guess, let's see, I think I'm I think I've seen this paper before.

C

Let me share my screen.

C

The link is in your chat if you want to follow along.

C

So this is a major challenge and analyzing animal behaviors to discover some underlying simplicity and complex motor actions. So, yes, this is ready. Analyze, the behavior or the movement behavior of C elegans were freshly reconstruct equations emotions from dynamics in this space I think this yeah.

C

Exactly yeah, this is a paper that people have been have passed around in the open. One group like open worm as a group, I called the movement, validation and so they've actually done some really good work. Let me give you a link to their database.

C

You might find it interesting if you're think it's, it should be up, it was done for a while, but they have like a lot of movement down are there and they have like it's a movement database, but they have done some analysis on this. So the you know you might look over that site to see get a sense of what people were doing with like mu were movement. Yes, this is where they come up with the eigen worms.

C

So this is where they actually take the data and they boil it down to like you know they use a dimensionality reduction, and so you know there there are a lot of statistical techniques that we are talking about directly here, like dimensionality reduction and other things that are nevertheless actually quite useful. Aside from the hard core machine learning, algorithms, like things like effective, is there anything you want to say about this paper.

C

C

Yeah yeah, actually I, couldn't hear a part of that. Could you type it in my chat? We could I could improve it back to the group, because I was yeah ever broke up a lot, so yeah I mean this is this is a good example of the kind of thing that you know and again this isn't like machine learning, but like it's, you know something that is useful so like in the movement group.

C

They take a lot of microscopy images of more movement and you know then they analyze it and they use this kind of space. You know like a reference space for the movement see or like basically finding modes of movement now in C elegans. That's relatively easy only because the arms movements are stereotypical. So, like you know, if you're analyzing human dance, it might be a little bit harder because there's a lot of like improvisation and there's, like you know, a lot of different degrees of freedom, let C elegans, of course it's easier.

C

But then it's maybe not as easy to like really kind of confirm that you have modes. I mean we're just using the human visual system and we're saying yeah: we can see modes of movement, you know under a microscope, so I mean you know, that's that's what I don't want to have like a statistical model, but then you know this actually is a pretty good way to do this. So you know they use this term eigen worms, it's just you know finding.

C

You know, eigen eigen values and eigenvectors of you know coming from the these movement modes and then these for Jesse. He any he's interested in attractor basins in that they do actually use attractor models of movements, so they're actually able to create. You know different attractors for different postures of the worm, so there's some dynamical system stuff in here as well yeah. So let's see what what our Chet we have here in the chat: okay, yeah yeah, that's fine! Abraham! Thank you for the paper.

C

I was very interesting so and then Jesse says definitely interested in the cognitive modeling aspects. These slides or videos will be online somewhere missing, chunks, no and then yeah I'll post, the video and the slides else. I actually posted on slack for you Jesse. Do you have anything else to add or any thoughts.

C

Okay, yeah, that's that's fine, and if you have any questions over the course of a week could be like insights, you can post them in slack. So thanks for attending okay, it's we're coming up towards the top of the hour. So next week we'll see what I can get someone else. Does anyone want to contribute something? Next week you have any ideas that you want, even even if it's just like a couple slides and bring up an issue that you want to talk about. We can talk about it.

C

Okay, so you know you don't have to. Let me know right now, but let me know before next week we could. You know I can send out an email a couple days before the meeting and we can yeah yeah. We can find a topic. I mean there are a lot of topics we can talk about.

C

You know it doesn't have to be super polished, but just just something to ignite some conversation, because I think it's useful DUP did. These were very good conversations today, so I'm glad that we got some. You know discuss this a little bit and okay. So uh let me know if you have any questions.