DevoWorm Lab Meetings, 19 Apr 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DevoWorm (2021, Meeting 15): Graph/Cellular Automata NNs, Diatom Cognition, Origins of the Embryo

Description

Graph Neural Networks, Cellular Automata rule discovery using Neural Networks, Diatom Non-neuron Cognition, Neuromatch Academy, Origins of Developmental Systems and Ecological Niches. Attendees: R Tharun Gowda, Debojyoti Chakraborty, Ujjwal Singh, Jesse Parent, Richard Gordon, Susan Crawford-Young, Krishna Katyal, Vrutik Rabadia, Bradly Alicea, Mainak Deb, Vash Yadi, Abhishek Tiwari, Shruti Raj Vansh Singh, and Akshay Nair

A

Hello, hello, hurry.

B

Don't know why it's fuzzy! Oh.

A

B

I can just turn it off too, but cause it's just not working right.

A

Oh well, this yeah the sun. I guess I don't know if.

B

Yeah there's a window over here, but and that's it bother it. I don't know.

B

Over there yeah.

A

Okay, well welcome to the meeting everyone uh so far. It probably be people coming in the next few minutes. So um yeah is it? Does anyone have anything they'd like to present this week? I know not everyone's here, but I know a couple. People said they might want to present something.

C

A

D

C

Hi, um actually I don't have anything to present right.

E

Now because I just started with my gui work, which I promise.

C

So I I figured out a way to you.

C

Designing a ui, which I I mentioned in my proposal like.

F

C

But I I I'm thinking of you know, making it more interactive and maybe uh maybe more uh better for like.

F

How, like the rolling images and all that takes a bit of time so.

C

I'm just like you know, playing around with the front end right now, uh yeah. So I just.

F

Initialized the word I actually wanted to ask bradley.

C

I could move those words to the demo worm gator repository, or should I just keep it with my own and share you or include you as a.

D

Contributor or something.

A

If you want to issue uh a pull request, that would be great what what um what project does it fit into most closely.

G

Smarter on the telephones.

A

Yeah, what what project does it uh fit into most closely? I'm trying to figure out what repo you could push it to.

C

uh Well, it was this like the accelerator embryo thing, so I have to I'm actually making the front end like the gui.

H

I promised in my proposal, so I thought why not start the work right now.

C

Because I'm having lots of time right now and my uh university exam should be coming up later, so I would get really busy so.

C

F

Just initialized initialized work.

C

F

If I get into some stages later.

C

On like right now, loading images is done and if I, if I uh get into the distortion, part labeling adaptation process,.

F

Or something we can even go.

C

A

Yeah, I think that would be great. I I don't know where. To put it right now, I'd have to kind of figure it out. Okay, yeah! That's.

C

A

Totally fine, I I hopefully.

C

By the next week, okay, yeah.

A

All right that sounds great. Thank you for the update. Yeah um we've got hello, mynak and josh.

A

I

I'm saying that is shining.

A

I

I guess she is sitting in front of oh yeah. I guess.

A

So does it any anyone have anything to present. I know some people came in since I asked before.

I

A

B

I'm just going to turn off my cameras.

A

D

A

D

A

A

A

Is it visible now uh not yet it's okay! There we go there. We.

A

I

So hello, everyone I'm going to present from now.

F

I

Going to present four tutorials.

F

G

You can say that.

I

uh Articles kind of stuff, uh where I'm going to debate how deep learning can be used for life sciences, apart from you.

F

I

H

Other domains that are not.

G

uh Specifically, computer science.

I

H

I'll be pushing the.

I

Code in github very soon, so this first thing is, as you know, that diseases are the oldest cause of human suffering from the dawn of human civilization. Women have been suffering from pathogens, cancers and neurological conditions. So one of the greatest achievement of uh last.

F

Two century has been.

I

The development of affected treatment for many diseases and the uh the problem I'm going to address today is regarding is regarding the solubility of a chemical compound and it's very important property for a drug, because, if a drug is.

F

Not able to you know it's not soluble enough.

I

F

I

Able to get mixed in the bloodstream of a patient and will not have its effects. So uh I I was using a deep chem. It's in library that usually.

F

Does most of the heavy working tasks.

I

For uh special chemistry based class, so the first thing that we need is a data that measures.

A

The solubility for real molecules.

I

A

The core components of deep.

I

Temp is molecular net, it's a diverse collection of chemical and molecular data sets. So we use, then this valuable data. So this is uh this is a screenshot.

F

I

You can see that here we have a compound name. We have oh prediction, that's uh you know how the solubility is. Then we have molecular weight. We have a polarity.

F

I

F

I

uh Other things, so uh one doesn't, you know, uh really needs to have very good grasp of chemistry. Working on these data sets, because most of the thing that we are interested in we are in this thing, is you know, getting an x and predicting and why, having a.

F

Domain specific knowledge is, you know good, but.

I

Not very explicit so- and these are the uh okay, so one more thing that uh for images we use uh convolution neural networks, but uh since uh this data set is not explicitly uh image based, uh it's not exactly tabular, also because uh all the molecules kind of form a graph like structure.

I

So uh I will be using a graph and resolution networks.

G

I

Like the foreign image, we have different uh pixel values and splitting them we can, you know, uh get all the convolution functions in data max cooling, average.

I

A network to get.

F

A hidden representation of a node, for example. Let us control a red node here.

I

One simple solution of glass: convolution operation takes the average value of the node features of a particular node, along with its neighbor, and that is a different.

F

From image data.

I

The neighbor of the node are.

J

Unknotted and variable.

I

J

In case of a regular 2d.

I

In a regular image.

F

That's not the case because.

I

They are ordered. You can see that uh when we have the 2d convolution there, you know they can.

F

Be represented in case of a definite matrix.

I

F

I

Them have a different uh gap between them. Okay, so this is how cnn works uh most probably every one of you would be knowing its uh usage. So I'll, not explain that uh and for graph resolution network uh our image can be somehow uh its dimension can be increased. uh It's not you know making it into 3d, but it's a way of kindly uh visualizing it. So we have all this.

F

I

F

This chemical compounds can be treated as a node. Then we have a.

I

F

Value for a particular molecule.

I

So that, uh after uh using the pre uh defined uh data uh after using the predefined uh uh graph, convolution.

F

Network the results I got without much of.

I

uh Tweaking into the network on four.

F

I

Was a root mean sphere uh value of 1.7, and since you can see that there is difference between train set.

F

I

This usually happens in mostly.

F

I

Neural network- and it happens because uh the test data is usually not is not shown to the model, so it performs a little bit less accurate shown test data, and this.

D

Is the form of root mean.

I

G

I

G

Are the references I highly encourage.

I

uh Everyone to you know, go through them that deep, deep chem library is really wonderful and most of the things.

F

Are predefined in it, so a person who is a chemistry person who doesn't have much of uh programming.

I

Knowledge can he each will easily do stuff on it without much of his timing based in learning programming, and these are the two papers that were really helpful, so these are also available online. So that's it any questions.

A

Yeah, well, actually, why don't you bring your slides back up um yeah any questions, uh because I have a few, so I okay, so if you can go back to maybe let's see where do you introduce uh graph convolution like five, okay, seven, okay! uh So this so the difference between graph convolution and like say, convolution in a regular neural network, is what you're taking like uh a structured graph.

A

Yes and you're doing this a similar thing to what you would do in convolution in 2d.

I

Yes, yes, like for uh uh normally when you should record all of these, uh you know uh somehow equally listen there isn't. There is an order, for example, if I want to uh blue max ruling here, all I have to do is find the neighbors and every neighbor would be is somehow equal, but that's not the case.

F

For uh graph convolution, because there isn't an ordered uh thing going on here,.

I

F

uh For example, if I have.

I

To come in the layman, if I have to do some operation of this pixel or this uh node so.

G

There are chances that the closest.

I

One would be most uh probably affecting it, but I have the neighbor that is really distant, so the effect it would be having on that particular thing would be significantly less. So the order is lacking here and uh the weightless average.

A

Okay, uh could you go to the? I think the next slide, uh all right so now in the bottom, you have an example of a molecule, and then you have a graph.

A

So is that graph being calculated as sort of a representation of that molecule like? Is it a.

F

Spatial representation.

A

Or is it what what what does it describe.

A

Okay, so the molecules here on the left and then the next thing there is a four node network, so yeah. What does that represent about the molecule.

I

uh It's kind of an uh you can say: digital.

A

F

Of the network.

I

Of the molecule here and it's uh being represented into an uh graph, uh we are just modeling it in a digital way. So like like, when uh for image, we.

I

A

And then you have the structure where you're looking at it, but of course networks can have different structures.

A

uh So you don't have the lattice property where it has nearest neighbors at a regular uh like you would have, and then it has like it could have like two nearest neighbors. It could have 50. yeah yeah yeah because of the way that uh networks are graph. Networks are connected versus like a lattice, where you have like a a uniform distribution of neighbors so yeah. That's, that's! Actually an interesting property.

I

Lettuce can be guess uh according to me, not sure like, for example, if you have, uh if you are modeling.

F

Diamond or buckminster fullering.

I

F

Are very perfect kind of.

I

Solids so they.

E

I

Represented uh in form.

A

All right yeah, we, you know we the paper that we wrote on the anons and bnns. uh We kind of talk about like uh neural networks, and we talk about brain networks and one of the things about brain networks. Is they have this graph network structure and so.

I

Yeah better at tolerating the fault, because these cnns are very much. You know.

F

It doesn't need to.

I

F

Kind of a perfect lattice.

I

They can be the topology that they are providing is very uh kind of.

E

Decentralized and.

I

uh Kind of much more fault, tolerant.

A

I

A

I

A

So do they what's the performance? Look, like I mean, are they I'm not I'm not kept up with the literature on that, but do they do? Are they comparable to like a normal convolution architecture? In terms of, like I mean, have people done that comparison or.

I

Okay, uh I have not got.

H

Much into the art legislature.

F

I

But they are relatively a new term, because uh c aliens are being going from like uh after the invention of lx90 in 2012, uh cnn are being developed.

F

Religiously, but for graph neural network, it's quite a new term.

I

And it works started.

I

A

That's interesting yeah I mean I would expect it to, but I don't I mean I don't know how easy or hard it is to implement. I mean there's some things that are claimed to do really a lot better, but they're very hard to implement.

I

All right, these are uh I, these are just kind of you know. Data structures for representing things like a matrix can be done for an image, so a.

F

Graph is, we can say a much more.

I

Versatile data structure, so that's.

F

Just the manufacture manifestation how we.

I

Did it so uh if it's more versatile, it can be used, for you know a variety of tasks, but yeah a lot of work is going on.

D

I

I guess in north near future, maybe in coming years uh there'll be, you know, there'll be the new state of the art.

A

Yeah yeah: let's see.

I

These these have I, according to me, only one tenth office search is going on there. Most of the things are happening.

C

I

The you know, language models and cnns. These are comparatively very much new and unexplored yeah yeah. You don't pop that much.

A

Yeah, I think that's it. I know we don't talk too much. Well, we don't talk enough about networks in this group, but we do do network research uh and we're going to be uh at networks. 2021, so that'll be uh something I don't know like. We have an accepted talk at networks 2021 and it's a virtual event.

A

So if people want to attend, I can send the information out um and then I I actually uh and I'll show you in a little bit I put in a application for another workshop at the or uh they don't call it a workshop, but they call it like a parallel session at the event, so uh you know that'll be yeah, so I don't know we might talk more about networks soon. Yeah so uh did anyone else have anything they wanted to present. I don't know if.

D

K

How are you all right, how are you yeah so.

A

I also have something to say.

K

I was like working on cellulot and I thought I should have dated you know uh not much but yeah. I have something to say.

A

K

A

Share the screen yeah sure your screen.

E

K

K

Yeah uh so as a part of my learning about cellular.

B

K

I am currently working on growing new cellular automata, in which we have differentiated model of morphogenesis, uh so the idea over here is to use cellular automata, along with neural networks for art recreation.

K

Since last few weeks I have been working on celery automata, but now we'll be using cellular automata, along with the neural network for our creation for regeneration and further things that I'll tell you later so. The output of the project, which I am currently working on, it's not done first of all in progress, so the output is not just simple images, but they are like virtual orgasms. You can see that are similar to 11 body and they grow and respond to changes.

K

uh Let us examine right if it said that if it will cut into two halves, it is able to regenerate itself- or uh you know the parts which it has lost so that regeneration process of uh in real life, how to stimulate it in our virtual reality. That is what we are doing here um so yeah. First of all, this is the example of uh from a neural cell.

K

So the bottom diff is not simply a set of flames. It is a continuously growing and interactive model that interacts. That starts from a single seed, like from uh from nothing to a dog image.

K

F

It just saturates when it knows that it has reached the target image, and this is what.

K

It wants, but the biggest question here. The challenge which you are facing in the field is how to correctly tear a call when to build or what to build like starting from nothing and going to a dorm. It sounds weird, so how does the model and when you stop? That is what other questions so biology. Has all these questions figured out quite well. We have evolution samantha on his own that it has. If it is cut into two hearts, it has to develop two legs and not ten.

K

Otherwise it will be weird, but how to do this and make our computers know about this virtually and for that we have a watermelon neural network, as the name suggested. So in terms of cellular automata. The problem lies in finding the update rule like we know that in a cellular.

F

Matter what, basically, we have.

K

Is a single cell and its neighbors and we have to update the cells depending on this neighbor. So how do we know what update rule to be used and instead of finding the optimal update rule in biology? What we know is here here is that we can employ a neural algorithm that finds the rules over.

K

This is how it should be developed how the neighbors should go forward depending on it. So neural networking goes with several trials as some form of virtual evolution. That organisms go through, for example, for so many years and billions of years of evolution.

K

Image from a from a seed itself, so uh in the project.

F

K

Doing the focus.

F

On the cellular, automata model.

K

It's it basically.

F

Focuses on a cellular, automata model.

K

As a roadmap for efforts of identifying cell level rules, therefore, each cell, how each cell will develop and give rise to more complex and regenerative behavior as overall and the major task in hand right now, which I am currently working on, is to develop an update rule starting from a single cell, how to produce that multicellular pattern in a 2d grid where we have uh organisms that we generate.

K

So there are three actually four uh four four sub parts.

F

Of this project, the first one, which is the basic one.

K

Is how to simply train accelerometer to develop a target image from random ways from a single like, for example, as you can see from nothing, you have developed the image which is suitable for you, and you know the four uh images that you are seeing. They all are from four different models, uh several automatic models- and you can see it is not that any rule would work. Some of them are giving very bluntless results. It.

K

uh So yeah, this is what I am currently working on. uh The next thing that it would be is to use a dynamic system for getting this image, not a.

K

Part of the project would be and repeating the perspective field is like if we tilt the object, a bit, uh what difference that would cause to the output so yeah. These are the four experiments that I will be doing in the near future and the first one is something which I'm halfway with the for I mean I don't have very great results right now, but hopefully, next week, I'll get better with us, so yeah this project basically is a toy.

K

B

A toy hemogenesis and regeneration model.

K

B

Be used for understanding.

K

E

Control of regeneration.

K

So yeah- and this is what I've been doing recently.

A

Well, that's good! Thank you, uh so I I see if you go back to like slide, eight, maybe.

J

A

uh So that's so at the bottom. You have this image, that's being sort of generated from over time. So it's just kind of blurry yeah.

K

These are from four different uh models because they learned it in a different way. Oh.

A

Okay, so these are different models. Yeah these, these don't like are the these. Aren't like a it's, not a time series. It's like different.

K

Individually, there are like a single one, you can say when it begins, it starts from nothing and then it develops. uh So these are four four four that you can see now: they're, starting from scratch and they're forming different different things, because the update rules are different.

A

Okay, so the update right this.

K

Is one of the examples I need to improve it because uh it's not very good right now it bit of a more improvement and I'll probably show it to you next week a better version of this okay.

A

So the update rules change over time is the as they're learned. Okay and then you end up with refined rules, and then you end up with something that's sharper than before. Maybe more detail.

J

A

Okay, that's good um a bunch of things in the chat. Let's see uh abachak says, I'm sorry, I'm not able to present presentation as I'm not from biological background very difficult me to come up with models or features which will be useful as community and biology researchers.

A

I can nab a check you can. uh You know present on something you know we can help you through the examples. That's not a problem. I mean if you want to find something interesting you can present on it. I mean you know it's there's always some application to it. So uh I don't feel like intimidated by that. uh My knox is interesting. Presentation graph convolution seemed interesting. That was for krishna susan says I have a paper somewhere that says that molecules tend to accumulate on curves become more concentrated to concentration curves.

A

So uh if you could share that, that would be good uh debo. I don't know which presenter, which libraries did you use, um I'm not sure. Well I mean that could apply to both of you.

E

People like that.

K

Yeah, actually, I'm.

K

J

K

That is actually trying to replicate what happens in real life. We start from a single cell and then we multiply and multiply, multiply and form.

F

uh You know medicinal organism.

K

Kind of the same thing is happening here. What we have is, we have some uh set of rules and using the setup group we give them to a cnn and the cnn.

F

Tries to learn.

K

How we are you know, modeling this and after learning, all the updates, we give it a random noise and that image is converted into what we see right now. Different objects.

A

So yeah so dick had a comment here uh see my two two recent janus faced papers. So this is why I think we talked about the one of them last week in the meeting uh the janice face logic and he says feedback seems to be mostly local, maybe sometimes global in animal morphogenesis.

A

So you know feedback, there's feedback between the cells uh that that local feedback dominates, but sometimes it can be like farther out than local. So um I mean this is something that you know it'd be interesting to see. If the neural networks discover that sort of pattern or how you know how we can sort of assess it, uh deebo says, are you starting with a random seed.

J

A

Okay and yash says very interesting, so thank you, so that's it good. Thank you for the comments. um Thank you for crishna and surety for presenting. um We have one more comment in the chat here. uh Verdict seems to be a cool field.

K

It is, it is actually interesting, yeah, hopefully,.

F

K

Have more results next week it will seem more interesting and more people can come in and join us.

A

Yeah, oh dick says someone needs to rope in michael levin's, observation of bioelectric fields and morphogenesis yeah. We well. We talked about this last week too. I think uh uh mike 11, who I think, we've talked about a number of times in the group, he's done a lot of work on bioelectric effects and, of course we talked about the flatworm last week.

A

A

Yeah, so that's good, um so.

K

A

K

A

Oh you're, muted, too,.

A

There we go bit.

G

G

Okay, okay, look. Has anyone uh in your media, I saw you automatically. Has anyone ever.

F

G

To include electrical phenomena.

K

uh I don't know what that is, and if anyone has done that, I'm not very good.

G

In biology actually so, okay bioelectricity goes back over a century. It was observed in uh fish, eggs and things like that. Around 1900 there's a whole book about the 1940s on bioelectricity electricity, uh lionel jaffe discovered many phenomena uh having to do with bioelectricity, and the current protagonist of this stuff is michael levin and it's never integrated with anything else in morphogenesis that I know of okay. So you guys who have studied some electrical engineering, maybe should learn something about this.

A

K

G

The uh you know, for instance, lionel jaffee found that when the brain is starting to form, it's called a neural plate and it had, he found what he called large electric currents coming out of coming out of the neural plate and going back in in chick embryos, and now that's never been explained or incorporated into morphogenesis models.

G

As far as I know,.

D

G

Okay, so you know, michael lubin is discovering all sorts of interesting phenomena, but uh I've tried. I can't get him to try to incorporate it into anybody.

G

K

Yeah, that's good yeah. I can you can look into that stuff too.

G

Okay, so uh now, maybe maybe we should start putting together some papers on electrical phenomena and see if they can be brought in under the you know, into cellular autonomy and stuff like that, yeah.

E

That sounds interesting. We definitely.

K

E

Work in that direction.

D

G

uh It's could be certain multimodal stimulations.

G

Okay, because electric they're, invisible and so they're generally ignored. Actually when the cards are so small yeah.

A

We have some other comments. uh Deebo said eel, which me well, that's that's an electro sensation and I mean it's.

F

I guess it's a.

A

Similar thing, but it's a different scale um and actually eels have, uh but there are a lot of fishes actually that have electro sensation, that at the organismal level.

G

A

G

A

Those are, those are adult.

F

G

And uh you know adult electric yields what's going on here in development.

A

G

Cell communication.

A

um Susan says, calcium channels and cells are triggered by mechanical movements. So, as an observation, uh deebo says some sharks have electric sense: can they accept sense, electric changes in their surroundings? You mean sharks. I think that's. The whole reason for electro sensation is that they build this field up around the sense, organ and then they're able to receive, like you know, discontinuities in the field, so they create this electrical field and there's this sensation of like discontinuities.

A

You might have like prey going into the field or you might have some other organism come into the field and they can navigate that way. So yeah, it's it's. It's largely for sensation, for detecting things, a lot of fishes that live in uh you know brackish waters or very dark waters and say like the amazon.

A

They have electro sensation because they can't really see anything. So they have to have a way to see you know to sense their environment, um and then susan says there are calcium waves in cells and embryos. Yeah I mean that's, that's why we want to study uh electrical uh cell cell communication.

G

uh Bradley we have a whole issue of biosystems on uh on waves and cells and embryos.

F

It's coming up there. I think there might be some people going yeah yeah, that's the waves issue, yeah.

A

Yeah is that closed now or is it uh I think.

H

F

A

Couple weeks: okay, that's good, so we'll have to.

F

Look at it when it.

A

When it's done.

F

But there might, you know the special issues publish piecemeal, they might be depicted.

A

Okay, yeah yeah, we'll have to see what they I I looked over. I looked it over a little bit. There were a couple uh articles that looked really interesting, so we'll have to go over it.

G

Yeah there's a uh lionel jaffee, for instance, studied fucus eggs. Fucus is a seaweed that has very thin flat, leaves and the eggs develop in the leaves, and these eggs uh actually seem to have currents to go in and out through holes and establish polarity to the uh to the egg, and then one end becomes what's called a hold fast, which is where the new lj attaches to a rock or something okay.

G

So so you know there is some evidence from the direct relationship between electrical current and morphogenesis, but uh it hasn't been brought into multicellular organisms except for. uh But what michael is doing now right?

G

But you know he seems to be mostly on the discovery phase, not integrating with.

G

A

It's very yeah. It's very interesting work, though it's, uh but it yeah it's something that of course we could model it. You know just as well, I mean you know.

F

It's going to be.

A

Limited by our knowledge, but we can also play with the parameters.

G

A

G

It might be fun, for example, to take a cellular automata model, each cell as having a current going in and out of it, which can influence other cells.

G

So the current the advantage of a current is it's kind of a global phenomenon. It fades off with distance, but it's global and then it covers a large territory, and uh I don't know just play around with that and see what it gets with it. If you make cellular automation with these electrical rules,.

K

Definitely yeah.

G

Okay, I mean if each one is sort of a battery which generates a current, then influences the cell uh or influences other cells. You can see the possibility of maybe getting something out of this. I'm not sure what.

A

Yeah, so that's that's uh yeah. I think everyone else in the chat is basically agreeing. It sounds interesting and here's diffusion of electrical current.

A

D

A

Forward to seeing more on this uh looks really interesting and.

G

Yeah, maybe invite michael levin to give us lecture to the whole pool all right.

F

I'll see, I know, he's very busy so.

A

F

A

Know he's always very good.

F

Like I said, he's fine with this stuff he's having great fun yeah.

D

F

A

G

The worst, he can say is no, yes,.

D

A

Well, thanks for that, um I guess I'll talk about submissions next, we have some a little bit of news on that and then I wanted to talk about the digital basil area paper that were sort of uh you know. We promised to do this and it's the time is ticking down on this, so we need to figure out what to do so.

A

um Oh, let me share my screen actually.

A

So the submission sheet here uh this is, of course, our master sheet and we have a couple of deadlines coming up uh evolution 2021 again, if you want to submit to that the deadline is april. 30Th uh krishna had an idea, kill the winners. If that's still uh go, you know submit it or else you know we can find another venue and then there's this other paper that might be submitted as well. On my end, so the deadline for that is april 30. and anyone can submit something.

A

If you think it's a good fit um uh the diva learned paper, that's not been moved on very much, that's something we'll move on, maybe later or so many different deadlines. It's hard to keep up.

A

Then there's this growth form and the theory of deep learning uh abstract and I actually submitted this to net neurosatellite, which is so the netside conference, or, I guess, networks 2021.

A

They changed the name recently um that that main conference we have uh an abstract already accepted to that this embryo networks and connectomics abstract, and so um that's in the main session, and they also had a call for uh network neuroscience.

A

So I I submitted this to that satellite and I'm not sure if it'll get accepted, I don't know if it's even a good fit necessarily, but I mean I just put it in to see if, if we can get in and uh you.

D

Know I don't know there might be other.

A

Satellites coming up for that that conference, it's a pretty good conference, it's it's one of the best network conferences or the sort of premier network conference, and um you know, even if you just attend it's, I think it's a fairly reasonable registration rate. They've got a lot of good stuff on network. So if that's something you're interested in, I don't know if they're going to have any uh graph convolution papers, but uh that's you know they. It looks like that might end up being come that might end up getting integrated into that field. Who knows?

A

um But I think that's something that uh to keep an eye on uh so so that we have that satellite paper and then the main conference paper here or it's not a paper, it's an abstract. But uh the next thing is this: basilarian nonrenal cognition. So this is for this volume on the mathematics of diatoms- and this is uh it says due april 30th, but I actually got an extension on this for several weeks.

A

uh I negotiated that. So that's something that's coming up, but it's there's an extension on it. uh This is something needs to be worked on. uh It's not in the greatest of shape and I'll show you in a minute what I mean um we're still waiting on the open room, poster on a decision on that uh these things are done in green. So all these things are done.

A

We still have the boring billion work. uh This idea of a kindle book of a lot of the content from divorm ml, and then we have open uh conferences. Well, we have the n and b and n's abstract, which or actually it's a paper now so I think extended. Abstract was an old thing uh that we'll probably hear about that by the end of the month. I'm not really sure what their timeline is on that, but uh that's for a life 2021, uh it's another conference.

A

So then we have a couple other things that that are coming up: deadline wise. There is no near ips, which there's an abstract deadline of may 19th and a full paper deadline to may 26, and there are a lot of satellites to nurips as well. So if you want to, if you have an idea like in machine learning or deep learning, that might be something you can submit to um the mathematics of diva worm, that's something worm book uh potentially from room book, but we're still working on that.

A

uh There's a living machines conference, which I think is the 2021 version, has come out now and I think the deadline for submitting papers is may 30th on that. So this is a conference where you know if you're interested in things like you know, I guess the there. You know, there's an interest in robotics at the conference, but there's also an interest in other types of computational models. So that might be something we might do something for the society for developmental biology.

A

That deadline is passed, they uh that's more biological stuff, so I don't think that's something we're gonna make so we'll put that gray. So if you have any other things, you want to add to this feel free to send me things. I can put them on the list or put them on uh if you're.

A

If you want to put them on yourself, you can jesse sometimes updates this sheet. As well with opportunities, so um yeah.

G

Really uh two two things: uh the uh the stuff on testing donald williamson's ideas on symbiosis uh could be part of the boring goalie we'll see what comes on.

G

If there's, anyone who has had some training in uh blasting, uh comparing dnas and whatnot, you might want to get interest get into this.

G

What williamson proposed, but it's never properly tested, is that some weird organisms like tunicates and whatnot are actually uh fertilization of some animal by an entirely different sperm, and you end up with a compound animal, and uh this should show up in the dna, in some fashion uh we're trying to figure out how to do that.

G

Okay, so if you've got background in uh analysis of dna sequences and whatnot, uh you might want to join us on that.

A

G

Okay, so uh you know uh test the test of donald.

F

G

In symbiosis, or something like that, for example,.

A

You can put that in here.

A

There and then we'll just put like this.

G

Yeah and then the reason it might fit in the boring building, because one of the things we hope to get out of this is when these symbioses occur and it's just a hunch that most of them might have occurred during boring.

G

All right that sounds good. Okay. The other thing is I'm working right now on the computer simulation uh of uh a lattice model at the molecular level for how biotops move you know the the idea is to uh try to explain at a molecular level why diatoms don't move smoothly and moving tiny little very fast jerks and some mostly forward sometimes backwards.

G

And uh been having some discussions down with thomas harvich, uh where he thinks the diatoms can actually sense their environment through the through the race. So uh I don't. I try to figure out how to possibly incorporate that into the model, I'm not sure.

G

Okay, so uh you could put down a molecular level simulation of jerky motion of diatoms.

A

I know like yeah the uh yeah they actually well yeah. They actually like in moments of position like you know, it's like the there's, acceleration and and uh then there's like jerk and then there's snap, crackle and pop, which are higher movement dynamics.

A

G

They moved so fast. We we got a paper in the diatom motility book, which is in press right now.

G

Around 1979, uh linda edgar, observed that the motion had very high accelerations between movie frames. She was doing 10 frames per second, so we tried to do it with a much faster camera at 890 frames per. Second, we got exactly the same result: huge accelerations in between frames and I'll, try to explain that yeah, okay and where, right now, the 890 frames per second were about the limits of the camera and definitely.

F

At the limits of the microscope.

G

Because we had to develop all sorts of subpixel algorithms to to measure the motion so because the motion would have enormous accelerations between one frame and the next and between one pixel like so, we had to get soft pixel resolution to get the others for it.

A

Yeah yeah, that's yeah, so that that actually brings me to the um stuff that we're trying to do on uh basil area. So I I was in in the loop on the emails between you and thomas, so I was reading those I responded to thomas. I don't know if I got you on that ccdu, but I basically you know I expressed you, know uh interest or admiration for his uh kuromoto oscillator model.

A

So he's he's doing this promoto oscillator model to describe movement, and so there are a number of different uh ways. You can look at this, um the what we have in in this uh the thing I'm trying to work on here for the basilaria non-normal cognition. We have actually two kind of two papers that are kind of running in parallel here and the idea would be to integrate these, so one of them is based on the sort of psychophysical world model of uh motility, and this has uh this has a lot of.

A

If you want to look at this, uh it's on github, so I I put it here as sort of an open papers, but I think I'm gonna somehow close this up a little bit and I wanna work on it a little bit more. But if you wanna take a look at it, there is the link, um you know, there's some sort of organization to it. It's not very uh you know it's it's still in pretty uh in rough shape. So there's that and then there's also this idea of uh collective pattern generators.

A

So these are like you know in uh in um chordates. You have these or even in invertebrates. You have these central pattern, generators that generate uh oscillations, but these are neurons, and so in this case that we're arguing is that there's this collective pattern, generator, which is generated by collectives of cells, that you do kind of the same thing, and so this is something that is maybe running parallel to this psychophysical model, and this will have.

A

This is there's actually some a data set in using stick insects as a model and they're looking at synergistic central pattern. Generators. uh So you can, you know, that's that's one source of data you can use as a comparison.

A

Then these are some modeling things that were done a while back, I'm looking at just modeling the sinusoids at different phases, so the cells are moving but they're moving at different phases to one another. So there's this you know there are these different uh signals that are generated.

A

So all this basically is uh in very rough shape right now and it needs to be integrated and then expanded upon, and so I I'm just putting these links in here to describe it or so that people can look at it and maybe give some feedback.

A

uh I'm going to be working on this in the next couple weeks and then I'll come back to the group with a more polished sort of rough draft and then hopefully, people can maybe read through these and give some a little bit of you know feedback about, maybe things that we should add, or you know if they're things, I'm not really relying on data for this paper. I originally plan to have a data analysis, but I don't think we're in a position to do that. um You know there are different ways.

A

You could approach this, I'm not really sure uh you know. If there's like we don't I mean we don't really even have a simulation that we can build upon, but uh I think just a discussion of it or maybe you know I I would like to hear from people.

A

Maybe there's a way to sort of uh crack this nut in a very you know uh something that won't take like a year to develop, but I I'm just for right right now, I'm just kind of thinking about like writing, this up sort of based on what we have here, but just kind of expanding on this and integrating these two documents, and then that should be probably enough and then we'll see where we are with that.

A

um Other than that I mean it's just kind of uh you know well we'll see: okay, thanks as well, we'll read through them thoroughly. Once and again we have a lot of stuff, that's up and coming in the group. A lot of people were doing analyses and things like that, um and I talked to thomas about his data. So thomas is the person who generated a lot of the data, that's in the uh in the folders.

A

So a lot of me. Some of you have seen this data that you've maybe worked with it a little bit, but that's going to take a while to get that ramped up where we have a good analysis pipeline, something that we can actually publish um so yeah. That's that's where that stands.

G

Right does anyone uh have experience with the fast motor control?

G

For instance, uh tom thomas harvich could use he's got standard microscopes? You could use a system uh where you can do very fast through focus imaging of moving diatoms, so you can get them 3d.

G

Anybody got an experience with that kind of thing.

G

The typical motion is about 10 microns per second, so if you want to get them in 3d, especially the best lariat column, is you got to be fast? You gotta vibrate your focus fast.

G

Let's see what city lenses, okay, could you susan, do you know a commercial source for them? Maybe you could send that to uh thomas I found the piazza electric vibrator, but uh well it didn't call itself z lines.

B

Okay, well, I wrote a brief.

B

According to the author, who developed it they're easy to make. Oh that would.

B

G

Thomas is an engineer, so maybe he could do it.

B

Yeah they're, basically a lens that is at different focal points and and so where, um wherever you're in the lens, you get uh different depths.

G

You can synchronize that with a camera, so you can get images of different depths.

B

Yeah apparently it focuses on the different depths, so you only have to take one step to get the different. Oh, oh, okay, so you.

G

Could do it rapidly and stop motion of the best larry.

B

G

Okay, yeah that'd be useful because one of the problems is pictures have uh some cells in focus some and many out of focus and because there are slightly different depths, which makes uh cell segmentation very difficult. So if we get the sharp images, uh then uh it would be much easier to do it track them attract themselves.

B

Okay, I'll look up that paper as well: okay, thanks.

A

Yeah- and I think we had a comment- I don't remember who it was in their proposal who talked about using z, uh z, uh uh like the the depth of the you know, for uh going through different slices of the embryo, so the z-axis information, uh and so I don't know, I don't know who that was. But it wasn't.

J

A

J

uh I have a few images. I got an image data set in g-axis too for uh this embryo. I think data.

F

J

Which I have used to generate extra centroids from 3d images? The outputs are pretty good now, but I have to do some work upon because exactly algorithm doesn't works there. We have to think about z axis also, so I'm working upon it right now. So I can filter out the correct indexes of like correct centroids yeah.

G

That's good yeah that that that debt is probably from a dead embryo. The problem with best larry is you have to get that data in under a tenth of a second.

D

A

It's very challenging.

A

Yeah, okay, so let me uh I'm gonna go back and share my screen and I wanna make a couple announcements and then I'm gonna get into some of the papers and then we'll wrap up for today. So uh a couple announcements here, um the first is that the neuromatch academy is open for applications, so the neuromatch academy is something that happens in the summer happened. Let the inaugural one happen! Last year it was on computational neuroscience.

A

So if you're interested in learning more about neuroscience, this is a good place to go it's about a two or three week, uh intense period during july. So if you go to this, so this this repository has the course content. So you can look over what they're gonna cover.

A

Everything is open to the public, so people can reuse this in different ways, but they make it, but but the actual, if you apply for the academy, you get the benefit of interacting with people and doing a project and all this, so you can do the um you can apply to be either a student or a ta or a mentor.

A

I'm going to apply to be a mentor again this year for the academy, um and this is the so last year they just did the one on computational neuroscience this year, they're actually doing a second uh summer school in august on deep learning, and so um this is the they're they're because they found that there was a lot of demand for it last year, so they're going to do a lot of they're going to do another one on deep learning in august, and so I think that the person is running.

A

This has debugged this on their course at u-pen, which is the university of pennsylvania and they're doing a course on this now and they're kind of working out the details. So this is a new course, but it's been sort of worked through and debugged, but if you're interested- and I know a lot of you in this group- are interested in deep learning and doing a lot of deep learning research. So if you want to apply to be a ta for this, I would recommend checking this out. There's a. I can't remember.

A

If I have the link in the um I can get you the link, um I don't know if they have a link in this repository, but um anyways we'll we'll. If you're interested contact me and I can send you the link, uh the other.

H

Yeah go ahead, hey um I would I really uh encourage people who don't know about neuromax check it out, I'm I did. I did it last year. I was the intensive track. um Some people here may have done that too, but I did it last year. um There's a big need for t8. So if you are experienced with it both in the computational earthline course and the uh the deep learning course will need tas.

H

So please consider that if you have skills or know somebody that does um there are volunteer forms for that, uh but also no look. If you are a ta, you will get paid, they work very hard to make sure they get funded, probably atas, but you can also volunteer in a bunch of different ways from editing the material to outreach to communications to funding. There are many ways you can volunteer as well. I'm trying to volunteer the awareness. You know a little bit.

H

I did it last year, but I'm working on some now reading some of the editing uh this year. So there's a lot of different ways. You can view, and um there will be two courses and so kind of like july and august, and it will be the thing you do and I know I'm going to try to be involved as much as I can.

H

I have a lot of other public school in the summer and data already stuff this summer, but this is kind of one of the one things I'm most looking forward to because it's such a great community. It's such a positive atmosphere and you can do I don't quite know what they're going to do for the deep learning course.

H

I do know it'll be very good comrades kind of spearheading it, but but uh there will be a lot of opportunity to make a group project again in the computational neuroscience work as well, and I think that'll be a really cool way to either maybe flush out some things in this group or let me stop your incident or find other people's little differential work and do stuff there. uh I really everything is a great thing to do and if you really strap your time, you can't invest in like the intensive nature of the course.

H

I'm pretty sure they're offering a- and I don't know if you've heard anybody's family, but last year they offer kind of a general anybody can stop by and look at this and look at the lectures course, but look at less intensive tracking and more intensive track, and I think they're doing that again this year. That.

A

I'm not familiar with, but if they've changed the last year that they did do the uh observer track.

F

A

I mean they probably do that again this year. I imagine just because.

H

A

H

They're really focused on outreach and getting getting people access to the material and trying to get it very inclusive and very progressive what they're doing so, if you're interested in like educational processes and sort of what you know, the future of some learning uh stuff and conferences and technology and how to put together, I really.

H

Appreciate I I've had a great experience there, so that's why I'm trying to wrap it as much as I can, but definitely reach out to bradley or to me. If you have any questions and um hope to see you, there.

A

uh The other thing I'd like to point out is that we have. uh This is the first day of the uh incf assembly, and so this is a meeting of people across incf, and so, as you know, we're involved in incf quite a bit, and this.

F

A

uh The talk the diva learn talk. This is now up on our youtube channel. So if you're interested in um seeing this go to the youtube channel, we already have 12 views. It looks like so you can go there and check it out and we have um it's. You know it's about. 15 minutes, yeah, 13 minutes and it basically goes over the devil, learn platform and some of the things that we've done and I think they're going to be pretty interested.

A

uh It's gonna, be you know it's it's online. So the way it works is if you're registered. You can go in and and uh chat about it um in a online portal, and if it this were a physical conference. They'd have like booths and you'd do your demo in a booth, but this is sort of the same thing, but it's just virtual, so hopefully we'll get some feedback on that. um So the final thing I wanted to talk about today are the papers.

A

So we have a couple papers. I don't want to get into too many. I found a nice yesterday. I found a nice image of a zebra fish here. uh This is a reconstruction of a zebra official larval brain using 2 000 individually, labeled neurons.

A

So this is uh an image I can't do this directly, but I have the I had the file in here. I wanted to show the animation, but it won't. Let me play it, so this is a animation where they've modeled, all these neurons and some connections and then they're rotating it in space.

A

So uh that's kind of an interesting thing. um I was gonna point out to this book. This is a book that I've been working with to to kind of give. uh This is an older book on anal organisms and neurobiology. It's kind of interesting in light of the things that we're doing the basal area, it kind of goes through it's an older book, but it kind of goes through a lot of different model systems, and so this is something oh, this isn't the right.

A

One though um this is the con yeah the table of contents, so kind of talks about like a neural systems and protozoans and sort of the approach to neurobiology in microorganisms cybernetics in the behavior microorganisms membrane, potential and behavior.

A

So this is like you know: they're, not uh they're, actually, the sort of electrical potential that we've been talking about, not really uh bona fide neurons with action potentials, but nevertheless they have membrane potentials and that plays a role in this type of behavior, uh contractility of muscle cells and non-muscular contractile cells. So there's a lot of information in this book and I've been exploring it trying to get insights.

A

So you know, maybe once I get the paper in better shape, we can talk about some of these points.

A

If you're interested in this book, I can, I can actually put it on the um on the slack to see if you know, maybe that's something that people want to learn about, um and I actually I wanted to talk a little bit, go back to this topic of development of early life and uh talk about some of the newer papers, so this is actually kind of combining the boring billion with um embryogenesis and developmental biology.

A

So this is the first paper I'm going to talk about. Is this current biology paper on a possible billion-year-old holozoin with differentiated multicellularity? So what this is is it's. uh This is an in brief strother at all describes life cycle morphogenesis in a new billion-year-old microfossil, which may provide clues to the evolutionary roots of embryonic development in animals.

A

So uh several weeks ago we talked about the doshanto embryo, which is this embryo. They found in china from a plant, and so they they show this this embryo. They call it sort of one of the first embryos and they show like in the fossil record this embryo, and it's kind of amazing that you can get that kind of preservation. But you can, in this assemblage that they have.

H

A

This is another example: this isn't a protes, a billion-year-old protist, so protists branched off from plants and animals at about the same time it was around that billion year mark uh we talked we had that.

A

I don't know if you remember the figure that I showed where we had the boring billion, and then uh you had an oxygenation event at the beginning and at the end and at the end of the boring billion with the toxin second oxygenation event, you get this diversification of like plants, animals protists all at about the same time, and so this is one of these organisms. So this is a micros, a multicellular microfossil by cell embrasiori, so it exhibits two distinct cell types, um this 3d preservation and phosphate.

A

So this preserved different life cycle stages differential adhesion. We have contributed to cell segregation during morphogenesis.

A

If you see this, this is a naked stage. What they call the naked stage, then there's this process of differentiation in cell elongation, which leads to this stage here and then you see differential adhesion and cell migration, which gives you the cyst, which is differentiated around the edge, and then it's also has this naked stage in the center. So this actually looks like blastocyst in animal in animals. You know, if you look at a human embryo or a mouse embryo.

A

You see that kind of pattern where there's an outer edge and then an inner cell mass, so it. But you notice that these cells from the naked stage they're differentiating in in the middle of the mass and then they're migrating out to the edge and they're adhering to and to form this ring. So that's basically what they're observing so they call differentiated multicellularity.

A

So um again, this is another one of these papers. It's very paleontological, so there's a lot of uh jargon in it, but um so they- but let me take it from here- the mature form of bicellum consists of a solid spherical ball of tightly packed cells, a stereoblast enclosed in a monolayer of elongated sausage-shaped cells. However, two populations of naked stereo blast show mixed cell shapes, which we can infer to indicate incipient development of elongated cells that were migrating to the periphery of the cell mass.

A

So you saw that migration pattern in the figure these simple morphogenetic movements can be explained by differential cell cell adhesion.

A

So in fact the basic morphology of bicellum is a topologically similar experience to that of experimentally, produced cell masses.

A

Now we're shown to spontaneously segregate into two distinct domains based on differential adherent-based cell adhesion, so they know that they're models in cell biology, uh modern models which exhibit this type of behavior, so they're kind of drawing a parallel there, whether that's actually, what happened is not really clear, but that's what they're, basing this on?

A

The lack of the rigid cell walls in the stereo blast renders an algo affinity for bicillum selling unlikely. That means I guess that they're not uh yeah its overall morphology is more consistent with the holozone origin, so it doesn't or originate from algae. It originates this whole as a soloism or an origin. Unicellular holozoins are known today to form multicellular stages within complex life cycles, so they're basing this again an analogy with living organisms and what they're doing and they're finding a common ancestry there.

A

So the occurrence of such simple levels of transient multi, so we've already seen here, is consistent with a holism on affinity.

A

So they don't, they can't make a precise phylogenetic prediction or placement in the tree of life, but they're just pointing to these fossils as being- and this is where they found it in this area here, just so, you know it's real paleontology and then they show pictures of the bicellum brassieri in its mature form. Has this ring so the entire mature phenotype? Is this ring and inner cell mass? So that's, basically, what you have is the mature form and it's developing from like an undifferentiated set of cells into this mature form.

A

So they show examples here of uh this. You know they show a lot of microscopy here and then they have some resources. So that's that paper. uh The second one is this development development, developmental capacity of the and the early evolution of animals by douglas irwin, and this was something I think that was accepted recently.

A

A

So here I employ a recently introduced conceptual framework for novelty and individuation that distinguishes between potentiation novelty, innovation and adaptive adjustments to the edicarian cambrian radiation. So this is something that you know. This is something again from early evolution, so the early evolution of animals involved the introduction of genomic developmental morphological behavioral novelties, and this is what they call the individuation of new characters, and so this led to the construction, new ecological networks.

A

So they had these organisms that were undifferentiated and then they started de finding these different sort of niches or they found niches, but they did so by diversifying their phenotypes. So you know the cells would diversify functionally in some way and then they'd do something different from their neighbor and they could form these ecosystems, which would then uh allow you know. uh You know it allows ecological sort of differentiation.

A

You get a lot of. uh You know you get predation, you get other types of ecological relationships and so he's kind of linking uh developmental differentiation with this ecological differentiation and so the origin and early history of animals included changes in cellular gene regulation and development during an interval of environmental conditions, including low and highly variable oxygen levels, low nutrient levels and low productivity.

A

So there was, there were a lot of these ecological challenges, many more so than say the history of life since the end of this period- or you know, since maybe in the last billion years, where you had really big fluctuations in oxygen and in productivity and so a lot of the organisms that were alive, then you know they had to. You know, evolve according to those challenges. They had to meet these challenges, and so this is where you sort of get this emergence of.

A

um You know not only of of complex life but of ecological niches, and so the feedbacks and interactions between these three factors render testing alternate scenarios challenging and consequently many discussions of this radiation in the edicarian cambrian is focused on a single factor. So uh evolutionary developmental biologists have largely ignored environmental contexts, while geochemists and paleontologists often view environmental changes particularly increased levels of oxygen as generating an evolutionary opportunity to which animals responded.

A

But they don't talk about the developmental aspect of that as much. So this is the this is a framework that sort of unifies these two things and so kind of goes through a lot of the different details of this um talking about developmental capacity of the metazoan, which is the the sort of the common ancestral animal and so the origin, in early diversification of metazoans, required the generation of new mechanisms to regulate multicellular reactions, and this include adhesion molecules, basement membranes and transcription factors.

A

So you get all these things that have to emerge in order to get these early metazoans and so he's kind of using this framework to explain a lot of this, and this leads to different cell types, of course, which then lead to these different issues in the environment. So um I think this is a pretty long paper. I don't know again this, but this gets to this. This larger problem, this larger issue of early life and development, and if you are interested in that um I would we should talk more about this.

A

I'm just continuing on this. This theme to see where we go or if we can you know, maybe incorporate this in some of our work as well. So I'll go back to the chat here. um Susan says there is a youtube about organism, called organisms cognition without a brain yeah. Is it a video or a channel? You could send a link along if you want um so sorry. I have to leave now have a nice week.

A

Everyone thank you as well for attending um susan says that it's a it's a science presentation, one in a youtube. Video just needs the fracking pattern. Jesse says, links to papers and I'll put the link in the chat here. Let me get the meeting link here.

A

If you need permissions, let me know.

A

Okay, so that's the link to the papers. um This paper sounds super interesting to me yeah. This is definitely something that um it's not something I think people a lot of people have talked about too much. I mean they're papers on it, but it's not. You know it's a highly, maybe an abstract or a sort of a obscure area. So um I am interested in this paper as well.

A

Okay, so it's in the folder there um and then uh yeah. So that's! I think, that's it for that. So uh if you stuck it out this long thanks for attending, I mean thanks to everyone for attending, but you get a special uh treat here.

F

And I don't have anything.

A

Actually for a treat so, but thank you for uh for being at the meeting here this week. If you have anything uh anything you know any any papers, you want me to review, you can send them along in an email or on slack, and if you want to do anything or you know, if you have any observations over the course of the week or any questions, you can communicate those via slack or email as well.

B

I'll, send you an email with the um try to find that youtube.

C

B

And also the um microscope.

A

Okay, that sounds good.

H

All right, oh yeah, I I think something is slack.

H

Just a few questions about uh some things sticking out today in the meeting and also follow up along the paper.

A

That sounds good. All right. Thanks for attending have a good week. Bye, bye, bye,.

A