DevoWorm Lab Meetings, 8 Mar 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DevoWorm (2021, Meeting 9): DL for Cell Division and Diatom Colonies, Cellular Automata, Microscopy

Description

Deep Learning for Cell Division (ResNet) and Diatom Colonies (YOLO). Custom Cellular Automata for Pattern Formation, Cellular Reprogramming, and Autophagy. Ball microscopy, GSoC Onboarding, and Zebrafish goodies. Attendees: R Tharun Gowda, Assaf Wodeslavsky, Aayush Kumar, Yash Vadi, Krishna Katyal, Bradly Alicea, Mayukh Deb, Mainak Deb, Akshay Nair, Ujjwal Singh, and Shruti Raj Vansh Singh

A

Hello there in the next shape.

B

Hi hi hi. How are you.

C

I have some doubts regarding the paper.

A

Oh, I don't think she's joining today. She she said something about her computer being. uh You know she had to have it fixed, so I don't know, but I'm gonna I mean I can probably answer some of the questions too. So I have some. uh I have a picture of what she was talking about last week, so yeah hello, my knock and ayush.

A

Hello, um you're, muted. I think I don't know your audio isn't working somehow.

A

I still can't hear.

A

A

Okay, welcome to the meeting everyone uh so uh we'll probably have a couple more people come in. We always do there's someone right now, that's akshay!

A

um So today we're going to talk about um follow up on some things from previous meetings. uh We'll have so a couple people said they could present something I don't know.

A

uh I think the roon is wants to present something yeah yeah and my knock maybe has an update on some things.

A

Is that correct yeah.

A

Okay, so uh first, though, what do we and then we'll have some other things? I have some presentations and some papers and we'll go over our deadlines. If you're watching on youtube uh you'll be able to follow our catch-up as we go along here, hello, my yoke. uh So uh why don't? We uh have some of the new people introduce themselves? First, I think we have. I don't know I think thoren introduced himself a couple weeks ago and uh akshay and ayush. I don't know if they've introduced themselves yet so.

A

Oh, your audio is really.

A

Low, that's all the way off now, if you want to you, can type in something in the chat.

A

A

Why don't we have uh um yeah? Why don't? We have actually introduce himself first.

C

uh I actually gave him an introduction last week if uh I think I used this knew so I would say so. uh Actually I'm making.

C

um I'm currently undergoing my undergraduate studies in computer science and engineering. I am deeply interested in like working with open source.

A

Organizations and.

C

A

Organization I've been working with was.

C

Kde, uh it's actually an application oriented software not totally.

A

B

C

B

Is a new challenge for me.

C

um Yeah, that's it well good.

A

Thank you for going over that again, and uh my next is hello.

D

A

You're audible, now, okay, good, okay,.

C

There was some problem with my headphones.

A

C

Okay, okay, then, I start my introduction, hello, everyone, my name.

D

Is aishwar and I'm a final year undergrad at.

C

Kit and my research interest mainly lies at the intersection of deep learning and ai for health care, and previously I have worked at iit bombay and I did.

A

The micro microcontrollers.

C

And small cell arguments and microcontroller segmentation, and then after that I worked on the volumetric segmentation of the other glossy female muscles. So basically what we did was we developed a deep learning segmentation model, a 3d model and uh for carrying out uh wall method segmentation, and after that we incorporated a deep learning model into a fully and uh cg.

C

So basically I worked on that and once the cd model was developed, we uh delivered that product to ncbs. It's basically a national council of biological sciences, which is based in bangalore and uh since then, that work is over and currently working on. uh I'm currently working to improve the segmentation further by uh using coordinate convolutions as well as any adaptive, alias.

A

That sounds good. Yes, very good.

E

Bonjour, hello.

A

There's walls here so um so why don't we start with the presentation, so the roon? If you want to go first, maybe and then we'll go to um my knock. I think mine are still there.

A

F

D

So I'll be talking about data tracking using object detection. So the main motivation here is to um study the non-neural cognition of diatoms. That is given an input such as the intensity of light, the concentration of co2 and the nutrient concentrations of track. You want to study the responses of these diagrams, um as movement is the most apparent response, so it will be very much easier to study these, so we need to track the objects across multiple frames, finding their radio position with respect to other atoms and their velocity.

D

So an opposed could be using the opencv method, such as thresholding and masking to find the position, but these won't generalize well across as the diagram might change the light illumination and the resolution of images. So as we know that neural networks generalize well across multiple domains, so we'll be using a neural network based on detection.

D

So the most naive approach to direct objects will be to use a sliding window where we just take part of an image and run cnn classifiers through it to classify whether there's an image or not and another neural network to predict the coordinates and the width and height. But the immediate drawback will be that we need to do multiple passes over the image to get all the objects and when we pass a part of the image through the cnn, it is possible that the whole object is not present.

D

As you can see in this correct image, so one proposed approach was to use a yolo based object, learning model which just treats the classification model as a regression model. So basically it divides the images into a grid.

D

So where here you can see it's a three cross finger and assigns each grid to an object whose center lies inside the grid, so each grid makes a prediction of the probability of an object being inside the grid. The coordinates of the center of the object, uh the width and height and the class table so, for example, in this grid. If you're predict you, if you're trying to find whether there is a car or not so, there's not so the probability will be zero and the other predictions we won't care over.

D

But in this cell, as you can see, there is a car, uh but this is uh placed in both the slow rates, but the center of the car lies in this rightmost grid, so the discrete will be making the predictions of the center of the object, the width and height and the class.

D

So the yellow object by dividing the image into several threads, makes it possible to find almost all the objects through just a single pass. So we get a fixed number of bounding boxes prediction from a grid, so we don't make several um passes through the image, so it makes it much faster.

D

So this will be one of the main reasons that we choose yolo for object, detection because it is fast, so it can make real time. Predictions like four tractions frames per second or like 1 45 second, depending on the model and euro, is trained end to end. So the parameters are shared between the classifier and the bounding box predictor.

D

uh So if my ic has much more higher accuracy than two models which are having different uh uh models which predict on the classification and bonding boxes, um so these are some of the predictions that the model which we trained on the atom object. Detection.

D

um And certain uh drawbacks that we can observe in these images are: it fails to detect the objects which are very much close to each other uh because of the great uh the people, the brains, which can only detect a certain number of bounding boxes which are fixed, uh and so sometimes it might be possible that two or three uh atoms get predicted together into a one single boundary box.

D

And you can see this uh right bottom of the image and the preprocessing part also for preparing the data set, becomes much more difficult as these object. Detection models require bounding boxes to have their size parallel to x and y axis.

D

So we have to rotate the object in such a way that the bounding boxes, which we uh predict are always parallel to x and y axis, and this becomes much more time consuming and we can automate this as the image needs to be manually rotated and when, if you have a bunch of atoms which form a structure and they are bent, so it becomes uh highly impossible that we can rotate the image in such a way that the bonding boxes are.

D

So we cannot detect all the objects in a single pass, so what we can do before planning to do is to modify the algorithm in such a way that we can increase the number of meshes so that multiple uh bounding boxes can be predicted and also to increase the data set size so that it can generalize much better across multiple images and to overcome the pre-processing part of the images. We could rotate the images in all possible angles and predict the bounding boxes and switch those predictions together to get a single bounding box per object.

D

But this algorithm becomes become much more complex and become completely expensive. As you are rotating the objects in all possible images in all possible angles, and then uh many predictions.

D

And one more model which we can try is using instant segmentation where the model makes pixel device uh predictions of so it predicts the class of each pixel, but this preparing the data set for this becomes much more difficult and time consuming.

D

We have to draw the bond to draw the outline of each object, which becomes difficult when you're using mass attract fat, so we require some special equipment.

D

um Yeah, so that's it so any questions or suggestions.

A

Yeah that looks very good. Oh go ahead.

G

Thank you for this presentation. I actually had a question. Could you just go back to the slides? Please.

D

Yeah one second, okay, sure.

G

So could you go back to the slide, which was talking about the boundary boxes? I mean the portion where you were talking about the roof like here, they're ready, so you.

D

Have discussed.

G

About this thing, where you're trying to rotate and make every picture basically vertical, with respect to those um cells.

G

Yeah I get it so uh and if you rotate it, then the pounding box does not actually cover the thing itself. Am I like it becomes a diagonal sort of a thing then, so I think what you can do instead, like you talk about this, like, I think what I had an.

A

Alternative, in my mind, is that, like you, can use a segmentation map and then like use random.

G

Rotation like in this case, you cannot really augment the data set other like other than vertical flipping right or some transform, like translations like shifting in x and y plane. So instead, if you want to go really heavy on the augmentation, you should go for segmentation models.

G

Yes, you're telling me that, like segmenting it on a.

A

Manually, annotated.

G

Data set is really like, not good, but what happens is that, like we can just experiment at first and see how it performs on a test set.

D

Though, like preparing a data set, was a problem like to create the boundaries of the objects like we couldn't do it. I tried doing.

G

So what we can do instead is that you can use these boxes, maybe to.

H

G

See can you register those rotations, like you're rotating all of those to become vertical break yeah, you're rotating it using python? So you can reverse the rotations right.

D

Yeah, you can do it like in.

G

G

Data, what we can.

E

Do is we rotate.

G

It vertically, as you said, we generate the bounding boxes with your model and then we can use that bounding box to create the segmentation mask.

G

And then we can do all sorts of augmentation. I think that will.

G

And then we can rotate it back to whatever it was and then save.

A

G

Like with with pairs of images and segmentation marks, and then augment and do whatever with it, for training.

G

Segmentation is much more useful than bonnie boxes in general.

H

So basically, like oh, my gosh, trying to stage like like uh we have like uh explore the yellow, including boundary boxing approaches but ultimately like we have to go for the semantic segmentation, or instance, is also not going to work for us. We have to open 70 segmentation and the thing like manually annotating, the data set so basically like if you are making.

H

Oh, you dropped out as.

A

A

A

Can people hear me.

H

Hello, you can see like brown boxing is not going to like uh it's generating results, but we have.

A

To select particular images, this is the same problem that we have like.

H

Three years ago, when me and my friend are working on this project and we have, we have also rotated all those images and inverted it back, and we have 27 edition and that kind of thing. So, if you are going with bounding boxes and you are trying to rotate it.

H

H

It is not giving us very good results, so what you have to do is that it is a tedious task, I know, but you have to like uh label images like using different uh labeling tools. I know like uh india, those labeling tools. You still have to rotate the image because, like they also support only like x.

E

H

And they don't support the process, so that is the one we like so for anyone. If anyone is going to work for this project.

H

uh In this project, so, like you have to adapt images very good once you have very good like this did not require like a very huge annotation, like I think 230 minutes ago, which came with.

H

H

And I think many things have been improved in two three years, so we can also do that kind of thing. uh So that was my feedback like. Oh, I think right now you are on track.

H

I think we can get.

A

Yeah, that sounds good. uh Thank you uh usual and my knock and uh thirun for presenting um yeah, I didn't know. I have any comments I mean it looks good. I was gonna, maybe raise some of the comments that other people were raising, but I think we we got that uh taken care of, so we got some uh people. uh Josh uh is a new person in the slack uh he's here.

A

Welcome asaph is here as well, and surety is here so uh now it's time for my knock to present the thing he wanted to show. Can you share your screen.

G

A

G

Okay, can you see me yes, okay, so if something happens like if the screen goes off in the middle, can someone, let me know.

F

G

I made some issues. Yes, it looks good, okay, so I'll begin with the presentation. So this presentation is.

A

Basically, an extension of my conversation with bradley and mayo.

G

So I have been working with your cat sets a lot lately, so the thing is that I've been trying to reconstruct.

G

So I'll be exploring these I'll be exploring these avenues now so I'll.

G

Itself and the aim of it then I'll be speaking a bit about the training process itself and the details of it and then I'll, be speaking about testing.

A

The model and extracting inferences from the claim.

G

G

So this is uk cell. As you can see, the bottom is clearer, so this is basically an immortalized line of human t, lymphocyte cells and these cells are basically used to study, acute t-cell, leukemia and t cell signaling. So these cells they undergo a cell cycle which has three stages. The first stage is g1 stage. The second stage is g2 stage, so these processes right these cycle.

G

So this a cell cycle, it's actually a very gradual process, but the labels that we got in the data they're like they.

A

Are just labels like they.

G

See the labels that we got they actually do not capture the gradual nature of this cell setting process so we'll be exploring that soon.

G

Okay, so the network that I have it will be changed to basically classify the stage of the cell cycle from the cell images. So, as you can see here, that here are the three cells, the first, if, if the geostage, the second is the stage and the set and third we get the g2 stage.

G

So by looking at them, we don't really like. We can't really figure out which stage they are in actually but the, but the training can actually move on to it. Okay, so, coming to.

A

The pre-processing part.

G

So this part basically ensures that the data is suitable for the.

A

Data loaders.

G

So the first flaw that we found in the data is that the data was in balance, so it means that.

G

Basically, over sampled, the minority classes, so I basically over sampled the minority classes.

A

G

Like even number of samples from each class.

G

My plan was to train.

G

And and further and for the rest.

A

Of the models, it will be something very very similar, but.

G

With a different set of words, so yeah- and I also left out 2500 and three data points as a whole- set for future testing so moving on okay, so these are the changing images that we had. They were actual see. The actual images were 66 by 66. They were actually rgb.

G

G

So and I have also used heavy augmentation techniques using a regular albumin, so I guess that's. I guess that will help in the overall generalization of the one. So I have also used optima to automatically find.

A

The best training parameters.

G

And speaking of the network.

A

Itself, I opted for.

G

A resume with a custom linear list.

A

G

Because the lesson 58.

A

Has a lot more classes.

G

In the in its layers, but here we have only three so I have to make sure to make those changes and I actually changed the 10 instances of the model on 10 different.

A

Forms of data, as I had said,.

G

Before so moving on, okay, so after training, I use the 2503 holdout points for training for not for training to actually benchmark the performance of the hundreds. So while running inference on only one model.

A

We got an accuracy of 85.5.

G

Percentage and when I actually blended the influences from the 10 models.

G

Which is actually slightly more than the one before so as it turns off. When I blend the inferences, I get a bit more accuracy, which is good, so here are a few samples with their corresponding labels and yeah and the directions. So these are correct, of course, so let me go on to the next slide.

G

Okay, so dimensionality direction using gmap. So this part is a bit complicated.

G

So each image is basically and it's basically a vector which has the dimensions of 3 by 32 by 32. It's a very high dimensional vector right. So these.

B

Vectors can be reduced into only two dimensions.

G

So in this plot to the right, every point actually shows it represents one single point in the test: dataset that we had, which consisted of three data points. So each point here represents an image from the test cells and the color represents the predicted class.

G

So the special thing about it is that okay I'll test that on the next slide, so yeah, okay. So these images right they were.

A

Classified and these colors represent the classes.

G

So, as you can see that these dots, they merge into one another like like these vectors, these images they these classes actually.

A

Merge into one another there's no solid.

G

A

No abrupt line between the classes, they.

G

Are like it kind.

A

Of represents the gradual nature of the process that we.

G

Were talking about before- and there is also.

A

Considerable variation.

G

In this blue segment, which shows that the g1 plus images they have a lot of variations.

B

G

Them so this large.

A

Spread actually.

G

Represents that so another special thing is that at no point, while training that see at no point, we specified the order of the processes.

G

And the sequence in which it undergoes these states so uh like the model itself, came up with the conclusion that this s phase comes in the middle of the g1 and the g2 phase, so that itself is pretty spectacular actually so like. uh It could be also like functionalities like this.

E

Can also be expanded into other fields like some other cells or some.

A

Other organisms.

G

G

A

Really interesting to have these.

G

Reconstruction.

A

G

From uh trained union efforts and having the ability to know uh the stages in the process, just by.

C

Training the model.

G

B

I guess so that's.

G

All from my side and I'd love to hear some feedback or questions that you guys might have.

A

Thank you very much. That was good. uh I had a well actually could you bring the slides back up uh just reassure your screen yeah, because I wanted to look at like make. Maybe some points about it, so I, like the umap analysis, he showed me that in the slack a while a couple days ago, so this is the umap and of course we talked about this in a prior meeting like the how umap works.

A

So I refer you back to that presentation and they have a new version of this out now. So, if you're interested in methods, it's a good place to go but um the so. This is basically a dimensionality reduction of your different, your cells and their different phases and how they're sort of related to one another. So they overlap.

A

uh So I mean the basic cl. uh The basic logic here is that if they're separable groups they'll be like clusters and if they're not separable, necessarily they'll blend together- and this is sort of because this process is, you know uh it's the same cell. It's just going through these stages and you can identify it using the phenotype.

A

Then you get this blending across this arc that you see here. I don't know if there's any significance to this arc, but it's definitely.

G

Actually, really it's not really significant about this art. What we can cut. What we can improve from this as a rule of thumb is that if, if there are two different points, these two different points actually refer to different images. So that's the basic thing, but this art that has that really has no.

A

But this is the this is basically the important part is that they blend together. uh So I mean you know when you do when you look at any one image of cells unless you synchronize the cells in the culture wherever you're looking at them, they usually have like cells in many different stages across the image. So, like one image would be like one snapshot in time of a cell culture and those cells.

A

Unless you synchronize them using some sort of like technique that people use, then you get like cells that are in different phases or in different transitional points, and so you know you're going to get like it's not going to be like you can have everything in g1 everything in g2 they're, going to.

B

A

So this is good yeah.

G

Yeah, so so what I want you to add to that. Is that see the labels that we had like the labeling procedure that they followed they actually they. They have basically assumed that, like under each label, every image is homogeneous, but after training, the model, the modulation comes through. The conclusion.

G

Right so listen! This is pretty interesting, actually, like. We usually assume.

A

That, under one clustering of the images, have to be completely.

G

Homogeneous, but this plot shows that by using these approaches, we can infer that we.

G

A

Yeah- and you know it's, uh it depends now. I don't know uh you didn't. One thing you didn't do is tell people what your cut cells were. I don't know if you know what what the I mean. Sometimes they give names to cell lines that are like kind of. I can't remember. I looked it up before and some sort of model.

A

You know we have model organisms, we also have model cell lines. So it's some sort of cell line. That's uh um oh there. We.

F

A

We got, we have this.

F

So this is the.

A

Bring that up yeah a mortalized line of human t, lymphocyte cells that are used to study acute cell leukemia, t cell signaling. Yes, so this is a model cell type. um You know it's something easy to culture, and so this is the, and so these cells are going to be. You know you can um you know.

A

Different cells are different growth rates, different phenotypes, and so uh I think it's good that we have like these model cell lines, because we can we kind of know how they behave and actually for like, if you're doing something like this, it's good because you can, uh you know, maybe build a you know, build a model where you can characterize what the cell looks like what things should look like in the cell see like here. You have the g2, the s and the g1 class and they all have like very distinct phenotypes.

A

um Sometimes some cells, don't necessarily have that. So I think that was a good choice for that yeah uh did we have any other questions about mayok or anyone.

G

It was only the g1 plus and like 27 percent for the s2 for the g2 and the s class. So it's it was actually every imbalance which had to be balanced to make sure that there's no inherent bias in the train mode.

D

So what I did is that I took.

G

G

H

To like induce some sort of biases in the dataset, so I would prefer steps.

G

H

G

If you're speaking of the force right yeah or for creating the force itself, I have used mksf.

G

Any questions or feedback.

A

Okay, yeah, I mean if, if people have feedback uh for my knock, we can talk about it on slack um but yeah, that's excellent, um excellent presentation and through excellent presentation. I know you guys have been working on this and this is great stuff. So why don't we follow up on this in the coming weeks and see how it goes? um I wanted to uh start going into some of the things I had prepared now, um and there are a bunch of things.

A

I know susan isn't here today and she had some computer problems, but she uh sent me some images or an image of her ball microscope. Okay, so we actually have something in the chat. Here I didn't look in the chat. um My yoke says the neural network learned the continuous representation order without the user explicitly feeding it into the model interesting so yeah. This was uh where we did it's.

A

It's sort of this sort of self uh self training, yeah, oh uh self-supervised type of thing going on here, so um yeah we'll follow up on this. So hello, krishna.

I

I

I

So see you in the next week so uh how many of you are for g, so here.

A

Hello, I think I think a lot of people are here for gsoc, probably well. You know yeah all the new people. I think.

I

Okay, so I hope you people got the google doc on the slides for g-shock uh bradley. uh Did you share them? The onboarding files.

A

Actually, I'm gonna talk about that in a minute uh yeah. We have some things to get yeah. We have some things to organize and send out so I'll be sending those things out this week. So let me see if I can pull them up here, um so we have all right. So this is the onboarding. So right now we have two onboarding documents and I'll put these in the make these more prominent in the slack.

A

The first is this thing that um krishna uh put together, and this is on open source uh sort of the open source motivations of our group and of summer of code. So this is a short uh powerpoint presentation and it kind of goes through what is open source?

A

Why should you contribute to open source and then a little bit about summer of code, the dead or the deadlines here? What a bit about diva worm uh want to contribute to evil learn. So this is a set of ways in which you can contribute. You know, of course you can just do your.

A

You know you can come to the meetings or you can do a project uh proposal where you can get selected and do a project, but there are other ways to contribute to diva, learn and diva worm, um and I'm gonna go over these slides and edit them a little bit before we do them. We put them up.

A

The other thing, of course, is the thing that my hook sent me, which is about the prod specifically about the projects. So we have the project descriptions up on um on neurostars and on github, but we have some actually a little bit of uh like what can I do before gsoc?

A

So you know, gsoc is roughly this application period which we have in the slide, so you have the deadlines there, but you know there are things you can do before you write your proposal, so you're going to be writing a proposal for gsac. What can you do before then? And so he has these things here listed which you can do before.

A

Gsoc starts for each project, that's very nice and then the skills and requirements here that you might want to highlight in your proposal or, if you really, this is something that you really specialize in. You might want to uh make sure everyone knows that or get you know, ramped up on it before uh g-suck and so there you know this is I'm gonna make these available here, there's also a general uh general faq that faq that uh he put together. So thank you for this. My oak and we'll probably be also we'll be putting these.

A

I think these materials up now I'll go through them and edit them a little bit and then make them publicly available in the slack channel. Okay, we have people coming into the document.

A

Okay, so that's good, um so yeah, and I think that should be good. I think for onboarding uh we'll have like you know the very specific open source part, the general part about the project. So I think that's good and thank you to both of you for putting that together, and so I wanted also wanted to point out here something susan sent me, and this is for the third project here. This is the ball microscope.

A

So if you can see my screen here, this is what you. So what this is is a microscope that images like a an embryo from all sides, and so what happens? Is you put your sample inside this ball? You put it up or actually, I think it's down here uh and well. This is the light source I think, but your sample will be loaded around here and then all of these these are different cameras. Here I don't want to comment.

A

These are different cameras here, hooked up to a capture device and they're, basically taking images from different angles. So you can see that they're coming up from below coming up from on top light source is coming in here, the samples in here, and so it's able to image simultaneously from all these different angles, and so now that we have all these images from different perspectives.

A

We can then take all those images and stitch them together into a sphere, so you can see that you can get. You can capture image information from various angles on the upper hemisphere of the embryo and then different angles of the lower hemisphere of the embryo, and then one of the the goals of digital microsphere is to be able to take those data that are being collected and stitch them together.

A

Now I think I I made it. I need the data that so susan has some data, that's available um that she made available on to me and I put them in a secure server, and I think I sent the link to one person who was interested uh yeah okay, um so that was actually um yeah.

C

Regarding that, like I have actually like, um I was actually thinking about monetizing pictures. So now it got clear like you were talking about stitching. So now it's totally fine, so um the sub sampling you know you sent like. Is it possible to create at least a demo for just the proposal?

C

Just for me to visualize at least.

A

uh I I think I might have some animated gifs, but that's about as far as I've gotten on it um just making.

G

A

But it's basically where what what happens with this, like uh each set of images, is so there's like. I think there should be like a folder of like about 12 or 18 images per embryo, and basically what happens is it's it's like you're taking a globe and you're spinning it along its axis yeah.

F

So you'll see that.

A

Those images are in sequence and if you stitch them together, it's kind of like it's rotating. Now you can do things with that. You can you don't need to build like an animation just from those images you can montage them or stitch them together in some way, I mean the way that you might want to think about.

A

This is like a geometric projection where you have, if you make a flat map on a surface, that flat map has been taken off of a globe right off of a sphere or off of a curved surface right and you're deforming that surface, but you don't want to deform it too much or else you'll be able to tell what the features are on the map. So you want to be able to. um You know, create like a.

F

You want you can.

A

Montage the image, the different images you want to take like the coordinate system, and you want to deform it enough so that it it flattens out or you know you can sort of see it clearly, but you don't also don't want to deform it too much. I mean that's, basically the idea. So in the proposal, I think we said that you know students can come up with their own solution to the problem and then it would be evaluated.

A

So I mean you know you have a lot of freedom in terms of how you want to approach it.

C

A

Yeah and one more thing like.

C

The subtitle you you.

C

Over time it becomes 14, then we will be uh exactly the fourth dimension is the time right. So I was thinking whether uh you send me the images of an instance, or is it just a random substrate.

A

I think it's a random sub sample. I think I I gave you like one folder that she gave me. I'm gonna have to go through it again and look and see exactly what's in there, but I think it's like just yeah just one instance. um You know there uh yeah. I would just uh we can go through it later and see what's in there, but um I think it's one instance.

A

I don't think it's like a time series or anything like that.

B

A

And so if people wanna, you know, people want access to this uh secure server that we have it on. Let me know, and I can send you the link um and uh you know it's like. I think that's a good way to distribute data. I don't want to make it public um yet so we'll just do it that way.

A

um So very good. Let's see we have some more things in the chat here. um Okay, usually asked akshay. Are you using 3d rendering models or trying it in blender, photoshop.

C

um Well, like I was thinking of using opencv, actually just switching it uh but yeah, I would give it a look like so far right now. I've got an idea about what the project should be like. After going through the research papers, and you clarified some dogs.

C

If it is not possible, then we would be you know moving to other alternatives. I would be going to it. I have a plan in mind, so I'm actually you know preparing my proposal also right now, yeah. I.

F

Hope bradley gets time to review it.

A

Yeah, I will review people's proposals when they have them. I can give you feedback on them, so just again drop a link in the slack on a private dm it'd, be fine, so uh yeah it let's go on here. um I have a lot of actually a lot of things. I was wanting to cover uh I'll, probably get through them. I'm gonna very.

E

Quickly, one quick thing I wanted to uh ask: okay, hello: yeah: go ahead, yeah uh yeah hi, so you want a few days back we're talking about the cellular automata thing, like I told you uh regarding different models that we can address. So I was wondering that if you could telling you in the future, you will might you might be presenting something or uh something like that. So lately I've been a little busy, so I could not do it a lot.

E

So if there are some resources that you would have in mind or something I would love to, uh you know have to cut them and get some idea about it.

A

Yeah we're going to talk about that in a couple of minutes, so we're going to okay.

G

We're going to go with that.

A

That's good yeah so, but I wanted to go over a couple things here before that um the first, my I'm not sharing my screen anymore, okay, uh so the first is just briefly to touch on the submissions, um so we have again some deadlines coming up, uh complement, which is march 26th, um the biosystem special issue, the periodicity in the embryo. That's been resubmitted already. uh Thank you uh bourgeois for your comments. They have been incorporated and jesse as well.

A

uh We had, uh I think it came out pretty well uh we'll be waiting for that decision pretty soon. So I I hope it gets accepted soon, but we'll see um this uh incf neuroinformatics assembly is coming up in april and the deadline for submissions for that is march 31st and they want a 1500 word or 1500 character, uh abstract, and so I've actually prepared something for that here. uh This is on the divo learn platform, so I I think I mentioned this briefly last time. We're going to walk through this devil learn platform.

A

This is the presentation that I we would give. I would take the lead on it and just kind of go over the platform, talk about the different components and then finally discuss a little bit about future development in terms of these epistemological directories, which you know um would be something that we would go over a little bit before hand in the group. I'll probably give a presentation on that and then also further development of some of the software.

A

So I'm going to submit this this, I can send a link to this for people to edit or to look over before I submit it, but then we can submit it and hopefully it will be presenting in front of incf. I mentioned in here that part of the people learn sort of uh development has been sponsored by incf, so hopefully they say look at that and say: oh yeah. We need to include this in the symposium, so we'll we'll be uh revisiting.

A

This uh also, I think krishna had to rejoin, but I wanted to point out that uh his work that he's been doing we've been doing on a ns and bnns, which are the artificial neural networks and the biological neural networks.

A

We're doing a paper for uh the artificial life conference on that, so uh we're going along on that there's still a lot of stuff to do, but um this is something that is due probably next weekend, so we'll be kind of getting I'll be cranking through that this week- and um you know you know- hopefully, maybe maybe people want to give feedback on it. I can send it out later this week when it's almost complete people can read it over and look at it um and then uh I think that's it for the submissions.

A

I think there isn't too much else coming up very soon. Europe's uh actually is coming up in may the the deadlines for paper, so people want to submit something to nurips uh it's very competitive, but if you want to do that, that's that deadline's in may, so it's coming up soon. um So just just to keep that in mind and then the evolution abstracts are now open again and we can uh submit something there by april 30th.

A

So that's now I want to move on to the um the cellular automata stuff. So this is a sureties. uh She just asked me about this, so I'm gonna.

A

uh I have some uh resources here that I want to point out, and this is a very uh crude sort of introduction to it, but I wanted to show you what we've been doing, what we've done in the past in the group here and some of the stuff I've done in the past on it. So we have this paper and this is available on the uh divaworm.weebly.com website. That's the group website officially, and this is in the publications tab, uh it's morphozoic solar automata with nested neighborhoods.

A

So this is thomas portages, who I think, gave a presentation in the group last year he's been a. He was a more active member several years ago and he's come up with this technique for a cellular, automata called morphozoic, and so this is a model that uh he came up with. That is open source. So we have the code on our github if you're interested in working with it.

A

It's I think it's written in c, and so morphozoic is a cellular automata that generates patterns that you might see in an embryo, and so we have these. um You know this. uh It's a platform where you have not just the regular neighborhoods that you would see in a cellular automata but extended neighborhoods.

A

So if you're not familiar with what a cellular automata does, uh let me see here: it is it's basically something like this. You have these cells that you know focal cells that you want, that that are active, so in this case would be the middle cell, and this middle cell has neighbors and it's influenced by the activity of its neighbors. So this cell here this white one in the middle, is the focal cell, and so it's going to listen to its neighbors.

A

But what are its neighbors exactly uh so in some cases in what they call the von neumann neighborhood. You have these black cells, which are the neighbors which are the direct neighbors in the cardinal directions in a more neighborhood, it's every cell that touches the focal cell. So it's all these cells around it, this nine, uh these nine cells in black here and then finally, the extended moore neighborhood, which can go out maybe two or three layers away from the focal cell.

A

So the idea is that these these focal cells are influenced by the state, the collective state of all these neighbors, and this happens across the grid. So you have a grid where each automata or cell is influenced by its neighbors, and so you get these patterns that emerge because they're all active and they're all influencing each other in parallel and they're, basically producing these patterns across the grid.

A

uh You know, they're, you know simple pattern formation uh in some cases like the game of life, you have things like things that move across the screen, like gliders, where you have different patterns that emerge, and so there are a lot of ways you can classify these output patterns and so in morphozoic they use uh you use a neural network to classify them.

A

There's a lot of detailed amorphozoic in in under the hood in terms of what the focal cell does. You know there are distributions uh that you can analyze within the cells, um and so you know there's a lot of information in there.

A

But basically the idea is that you have this, uh these sort of multi-layered networks or these multi-layered neighborhoods that um determine what the state of the focal cell is so, um and you can produce things like this, where you have this grid in there, these cells and they're all interacting and they form coherent patterns, and sometimes you can actually analyze these patterns as different the execution of different rules.

A

So each focal pattern will execute a rule over time based on the state of its neighbors and so it'll be like something like if a majority of my neighbors are off. I turn off if the majority my neighbors are on, I turn on, and so just with those rules being implemented, you can get patterns like this, so this is rule 30 right.

A

So this is rule 30 where a single cell is turned on in the top row and then next time step, all the cells are subjected to rule 30., two of them change state as a result leaving three cells turned on in the second row. So in this case the rule is to have these cells executing in sequence and then producing this pattern.

A

So there are ways you can do this that you know produce these patterns, and um stephen wolfram has a huge book on this and he's classified like he's identified at least 200 rules that these cellular automata follow, and you know they're all different. They all result in different shapes different patterns, but they all have like. We all know what the rules are and that went into creating them.

A

So this is like this is very analogous to pattern, formation and embryos. As we've seen in some of the papers we've talked about, and we you know the the good thing about using something like morphozoic is you can understand the rules that unfold? So if we compare that to something we're doing with image segmentation image segmentation, we're trying to discover the pattern of images right, that of a process- that's already taken place in this case, what we can do is we can actually sort of forward generate a pattern using certain rules.

A

So we know, for example, that certain rules have certain outcomes and and if we apply them in a computational environment, we can get these. You know we can simulate them and get different patterns and then understand you know which patterns are maybe replicable or you know, aren't just ephemeral things. So this is. um This is like what you can do with ca, um and so this is something that is a specialized package that we have in the group, but there are all sorts of different um packages you can use for this.

A

Everything from like a standard, cellular automata model to some very highly specialized model. And again I would recommend that you read through the paper and it gives you a lot of detail of how this uh technique works. So it's a very specialized type of ca technique. uh Oh we have so. This is referring to cell type densities. So there's a in morphogen neighborhoods.

A

You have these cell type densities that you can analyze as well as the state of the cell and the pattern on the grid. So there's a lot of potential here. I don't think it's really. I I think it's been sort of not developed enough. I know that uh tom has moved on to other things, but this is something that we might look good back into.

A

um Yeah, so this is- and this is the game of life which I was talking about before so the game of life is where you generate these patterns and they are mobile across the grid, and you see you know you basically set up the cellular automata and you run it and you can see these different things that look like life that are emerging on the grid and you you know they're different things in the game of life that you can see.

A

um You know you have gliders, you have other types of things that emerge and they collide and they reproduce, and so that's another way, that's another type of cellular automata.

A

So this is, uh I guess this is a very crude introduction to this, but you can have- um and this is sort of the proof of concept here, using an image image, restoration, and so um I would recommend that you read this and then you know, let's see if this is the one that I wanted. Okay, so this is like something where I was working on this for a project.

A

If I can come up with the poster, this was a pro star dynamics days. We're remodeling cellular reprogramming as a process, so the idea is these cells and culture.

A

If you transfect them with mrnas, they will affect gene expression which allows them to change their cellular state, so they go from like the skin cell type to the stem cells, and so the question is: how do you model that process in a way? That's uh allows us to understand it, and so you can use cellular automata and in this case, you're using something called a sort of a sliding neighborhood, which is where the neighborhood would actually deform, based on the nature of the process.

A

So you'd have these neighborhoods that where the cell would grow and then you know it would influence the or the neighboring states would influence a cell and it would eventually change so there's this. uh There are different ways you can set up these models. That's the point of this showing you this um and so the advantage of using a cellular automaton. Something like this is you can look at the spatial heterogeneity of a biological process.

A

So in this case, you have a bunch of cells that are aligned in a culture, and when you deliver this stimulus to the cell, it will behave in a way that you know the neighbor.

A

You know different cells will take up different amounts of the stimulus and then some of them will have different things going on inside of them that enable the reprogramming process, and so you end up with basically these colonies of transformed cells that change their state and some cells around the edge that don't change their state or die off, and so that's the idea. We want to understand that process better.

A

We also introduce a mathematical model to this, which is the uh fitsy nugumo model, which is an excitable cell model, so you can actually model this process as a model of excitability across space, and so each cell that takes up the the mrna that you're delivering to the cell and transformed successfully is excited and the cells that don't do. It are not excited, and this is a lot of parallels with uh neural activity. They use this model.

A

The fitzhum nagumo model in neural modeling, neural activity and neural excitability, and it's something we've been talking about to look uh for looking at the basilaria colonies as well, a model of excitable uh media, and so in this case it would be like you know, each cell if it's moving, there's an excitation there and a relaxation and you can model all that in a mathematical model, so another, oh, this is the same one.

A

um So this is this is the and then there's another poster where I think I advance this idea a little bit more. um This is where it's called dynamic: cellular encoding, dynamical cellular encodings. So this is a dynamical system where you have again these sliding neighborhoods and you have these inner cellular functions.

A

So you have the vocal cell and you have the neighbors and you're annotating the cells in a neighborhood and they're doing things across the cells they're, making comparisons so they're, not just making, like you know, applying simple binary rules, they're actually applying different rules and seeing if conditions are true or false, and so you can get you know you can use that type of information in a ca.

A

As well and you can incorporate that into the design and it'll result in different uh patterns, so it's not just a matter of having a grid where things turn on and off in parallel, you can apply all sorts of rules. You can even apply within each cell. You can apply things like the fitsum nugumo model or you can apply some sort of gene regulatory network model or you know you can do a number of things. There are a number of ways you can very innovative ways. You can approach cellular automata.

A

So that's I'm going to talk about there. um I think that's, like I said it's a very crude introduction to it, but I think it'll help people understand a little bit about uh and I'll, probably if, if shorty wants to do a presentation on it, she's welcome to otherwise I can do a short presentation on it in the coming weeks as well.

A

uh Okay uh has to leave. Thank you as well for attending and thank you akshay for attending. um Yes, so again, I'm gonna go over a couple more things before we end their meeting. I wanted to talk a little bit, maybe about some papers that uh I've found over the week so last week uh the twitter thread.

A

I put a twitter thread for the meeting up every week with the youtube video, and so you, if you're on twitter, you want to check that out and I have a I found a thing on a thing related to cellular automata last week on twitter, and it was this animation where using one cellular automata to feed another. And you can't see the uh I don't know if I have the gif here.

A

But basically this is where there's a cellular automata here and it's hitting a barrier and then it's transforming the output of this automata into some other visualization, and so there's this uh and I found out after seeing this. It was very cool. But I found out that after this there's something called cyclic cellular automata, and so these are things that uh recycle so there's a sort of trophic relationship uh with the cellular, autonomic and its environment. So uh there's this uh repo and cyclic solu or automata.

A

There's this animation uh and with I think, with the code as well, uh and so if you go back to last week's uh twitter uh thread on the meeting there is this uh thing that I put together here: cellular automata as a trophic system, and I call it computer fiji, which means that computational systems are eating one another or that they're being consumed. So this is the food line for cellular automata.

A

So that's I mean that's again, something we can revisit. um So I don't know how much I want to go into the different papers. I I did want to go into the zebrafish gastrulation thing that I found as well.

A

So this is a nice visualization of a zebrafish embryo undergoing gastrulation, and this is a zebrafish rock twitter feed here. This is where they have the transformation from uh a spherical, embryo and zebrafish, and it's and you have this uh gastrulation stage taking shape here where there's differentiation um along this axis, and so you can see in this image- and this was um this- is the person who took these images. So credit goes to this person um and then this is the actual gif which I have as an mp4.

A

So this is the process here, so you can see how it like starts at the top, and then the cells work their way down around in into a sphere you see like that and they and it becomes more defined, and then you get this axial differentiation right along here. This is the earl crest area here, so you can see how that works.

A

It's kind of coming together and there's the neural crest and there you go. So that's that's a nice visualization of that process, and I thought you know this is what I'm trying to strive for here in the meetings is to understand like what these things look like it's kind of hard to communicate. Sometimes so uh I think that's all I'm going to do for today. um I had a couple papers, but I think it's probably getting pretty late in the meeting, so um I I wanted to ask if people had any questions.

A

Actually let me go to the chat first. Okay, so we had mayoque says thank you, my knock and the rune for the presentations great work. I also encourage others to present their findings in the coming weeks. If possible. Yes, please do if you want to present anything. Let me know before the meeting on slack or by email.

A

We also have an email newsletter. I don't know how many of the new people are getting that, but let me know if you're not getting it and I can subscribe you. uh I just tells you about the next week's meeting um and do we have any other questions before we go.

E

No, I think it's, it was good. It was a nice meeting day.

A

Well, thank you uh and again, if you have anything you want to present in the coming weeks, let us know otherwise. Everyone have a good week.

A

Thank you for joining bye.

E

Bye have a good day. Everyone.