DevoWorm ML, 23 Oct 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DevoWorm ML: Week 8 (Computational Pareidolia, Blog post update)

Description

Eighth DevoWormML meeting, October 23. Attendees: Bradly Alicea, Richard Gordon, Jesse Parent, and Vinay Varma

A

Hello hi: are you that honey, okay.

B

B

When you make changes to the Google Doc, oh okay,.

C

Copy I, just kiddin me I, didn't get an email when you make changes, but no.

B

Maybe I should send those over cuz I think you missed some. My calmness, oh.

C

Yeah I'll look it over again.

B

Okay yeah: the problem is that Google duck doesn't keep a record of the changes or it's like you know,.

B

If you comment of mine, it just disappears.

D

Okay, but I get it I, think I get a copy of it.

D

I'll check those files, see I, see.

D

B

You seem to accept the reason and that's the problem.

B

D

Of them are questions rather than changes.

E

E

E

E

C

E

In the document yeah.

C

You can be.

C

Coming might get started and then, if they come up during the the proceedings- and you know we can let them in and they can catch up so welcome to today's deworm ml meeting I'm gonna present on something called computational pareidolia and then we'll talk about the pre-trained model, so Vinay we can go through that in a little bit more detail. I worked on it yesterday. Actually, you know pulling in basically flushing out the outline, and so we can go over that and then we can talk about what Vinay wants to add in more detail.

C

This is that. Well, let me share my screen.

C

Can you see my screen?

C

Yes, okay, so let's get started.

C

Okay, so today's topic is computational pareidolia, and so this is an example of a painting that was made by something called style transfer, which is where they take the style from one type of painting and they apply it to another image, and this is an example of sort of like a sort of perceptual trick that you're pulling them on the algorithm and we'll talk about that in a little bit.

C

So what is pareidolia? Well, it's a Greek word that means beside or beyond former image, so they come by and you know, beyond the image or beyond the form, and the definition, but according to Wikipedia is interpreting a vague stimulus is something known to the observer. So you see these examples here see we have Cookie Monster, where it's like a mouth. Video game is trashcan as a mouth, they put eyes on it and this power out what looks like a face, and this coffee cup has full foam in the shape of your face.

C

Our face is actually quite common in pareidolia, although it can be other things as well and even be auditory information such as songs that you interpret in some way that you, your brain, prefers, and so this is well documented in human perception, but I'd today, I'm going to talk about how machine learning algorithms can either simulate this or all victims.

C

So this is a good example. This isn't pareidolia necessarily, but it's an ambiguous image and so ambiguous images. Are these images where you have this thing here on the left? That is, you know it's a duck. It looks like a duck, but it also looks like a rabbit. So if you focus in on it the first time you'll see either a duck or a rabbit and then, if you focus in the second time you might see you know you might see the other image. So you see a doctor first time, then you see a rabbit.

C

So if you look at it kind of sideways, you can see the ears of the rabbit here or, if you look at it this way it's the bell of a duck, and so there are a lot of these ambiguous images. In this case, this is a Google cloud. Vision is actually able to interpret these emit these ambiguous images in either way, depending on how you orient the image.

C

So in this case the computer algorithm is interpreting it in two different ways, depending on the orientation and our variable systems do sort of the same thing, but we don't I mean we kind of understand that the visual system does it, but the computational system is replicated so I I wrote a blog post on this a couple years back, so this is the link to it and it's a short kind of a short post, but it has a comfortable articles in it that kind of go over these different.

C

You know maybe different ways that people have done. This and musings of people have made about this. It's not something! That's very common.

C

So in this they've actually taken this concept and turned it into a genetic algorithm that creates faces James. So in this case you have a, but you like, have a feel, a two-dimensional array with a bunch of geometric shapes and the stack together in different ways, and then you use this genetic algorithm to select for things that look more and more like a face.

C

So in this case we have a bunch of you, know: polygons that are overlapping and we select for the polygons overlaps looks like a face, and eventually we get to something that approximates the Mona Lisa. So again, we have something that isn't organized I mean it's just random polygons stacked on top one another with differentiating characteristics, different shape characteristics and they form this face.

C

Particular ah well you're, just I mean they don't necessarily form it. You know on purpose, it's just that. You have these shapes that are overlapping and you do rounds of selection on the different shapes that look more like a face. So it's like you're getting you're kind of bootstrapping this face into existence from a bunch of different shapes being presented, so I mean it takes a while for this to converge. You know, depending on the initial condition, but you get these different.

C

You know sets of polygons that overlap, and you know you get a like a population of these things in the first round and then you select the ones that kind of look more like a face. So you know if the this. The second image here shows that you have some formation of leek. Something looks like a face in hair, and then you select for that and then breeds a number another set of similar images again, and then you select for this and it breeds another set of animals like this.

C

Like this one missing, why didn't you come up with your face?

C

What do you mean my face?.

B

So somewhere in here is the that you're hunting for the Mona Lisa yeah.

C

Where is it well, yeah yeah, it's selecting on things that look. It comes more and more similar to the actual original image. Oh yeah yeah, okay,.

B

So me, in what step is the original image put in I? Think.

C

The first step I think using it from the first step and they're using it select.

B

Successive approximation boom goal yeah.

C

So there is a genetic algorithm and the fitness function looks like the Mona Lisa and then you you. Basically you take your first round of images, so you might have a population of a thousand images that look different and you say which ones look most like this original image and then you select those and then you produce a bunch of some other images to those you know you make a reproduction step and then you find the ones that are most closely related to the original image. And you do this.

E

Formation of a face day, yeah.

C

So in genetic algorithms you have a population of different images like a population of organisms, so you know there I, don't know what their population sizes for this I can't remember, but you know say it's like 10,000 images that have different. You know configurations of these polygons and you take the original image. You say this is the fitness function, so this is what I want to approximate and then yeah is that are closer to that. You know get it selected and though they don't don't so you get this and it takes like.

C

You know several hundred, maybe even thousands of iterations, to get something that looks like the image on the fur right.

E

C

It doesn't look exactly like a Mona Lisa, but our perceptual systems seeing it as that and the algorithm is seen as such. So again, this might just be: it's actually isn't a bunch and polygons we're just interpreting it as a face, so the neuroscience of pareidolia is a bit different than that, of course. So here's an example seeing Jesus in tennis, narrow and behavioral in Europe are lots of face. Pareidolia this is described in this article was an interaction between our face recognition and our visual perception abilities.

C

So in our brain we use our attentional system, uses these top-down mechanisms and these bottom-up mechanisms so attentional processing is you know, both bottom-up, so you're, finding random patterns in the environment and you're evaluating them, and then you have these top-down mechanisms, which are things like object, recognition face recognition, so you're taking these things that are random and you know you're applying some sort of you know mental model to them. So I know how do you recognize someone's face?

C

Well, some people argue that we have, you know neurons encode specific faces, but in general we can recognize a face more than we can recognize. Like some object, we have a stronger ability for that. We also have strong abilities for very familiar objects. Like you know, pencils or mugs things. You see every day versus maybe things you don't see every day, and so because of that we have this top-down mechanism that enforces some sort of order on all those sorts of things that we see in the environment so faces again.

C

Unlike other jaw, objects are processed just as a whole thing, so you look at something and you either see a face, or you know sometimes not a face. I gotta wet someone in the meeting here: hey Jessie, we were just in the middle of this talk, computational pareidolia, so I can go over it a little bit with you know with you afterwards, but welcome um all right. So let me go back to this slide.

C

So faces are processed holistically, so, like you see a face or you don't, and sometimes you falsely recognize things as faces as a result. So you look at a power outlet or an electrical outlet and it looks like a face or you might look at like some set of clouds and they look like a face, but they're not faces, and the reason for that, of course, is this top-down attentional mechanism, and then I mentioned before that you can also have things like auditory stimuli that exhibit the same phenomenon.

C

So I I, don't know if a lot of you've heard the thing about like misheard song lyrics. Are you listening song lyrics and you hear certain lyrics that are actually wrong. So that's another form of this pareidolia, where you're taking familiar things and you're sticking them into the into the ambiguous stimulus.

C

So, like you know, you might hear if the musician is like, you know, saying things and in in interspersed with the beat of the music, you might misinterpret some of the lyrics as things that are more words that are maybe more familiar to you. So that's the neuroscience of it, and this is not the neuroscience of it, but this is how you can replicate it in a computer. You just simply, you know, try to form something that looks like a face and then it'll produce something that looks like a face. But it's not really.

C

This isn't really a face. This is something that we recognize and for that matter something a computer can recognize as a face, because you know, if you put this into a face recognition algorithm, you might very well get a positive result.

C

So this is, of course, what we're talking about here when I say a false, positive and pareidolia is a false positive of like things like faces or shapes that you see in the environment. So here we have this Venn diagram, so you have true positives, which are this area here. This is, if you see a face- and it is a face- both negatives are things that are rejected as faces properly so things you see, and you say no, that's not a face. That's not a face. Both positives, of course sure this territory of pareidolia.

C

So this is where you see some pattern and it will take a face and then true negatives are things I.

B

Think there's some problem here in the very language that used to describe this because false word false has a very strong connotation.

C

B

B

Let me give you an example where this this mechanism is actually helpful. They have a false positive.

B

That is, you have to go back and evolution. Think about predator-prey relations. Without a jungle and I see a I think I see a tiger yeah. Ok, it may be a false positive. However, it is better than I run, even if it wasn't there yeah.

B

There's no harm if I don't run and it wasn't there. I bet like a bust, their own net and therefore, what's called a false positive, actually can be quite helpful. Yeah.

B

So I'm saying the word: the very use of the word false in describing these phenomena is a bias, not thinking about them. Yeah.

F

C

Was back to this part here where you know we talked about how faces are processed holistic? We, so you see something, that's familiar. You know it's good to see something. Maybe that's familiar versus. You know not being able to tell what something is in the environment. If it's like a snake, then in in the grass- and you see that, but it's not really a snake, then you benefit from that more than you don't. This is just kind of like the technical language of it in terms.

B

There's a value judgment made in using the word false, which may not be appropriate. Yeah.

C

Yeah yeah, it's like in you know we. This is what I kind of gave this talk, because I wanted to see. You know this is interesting. We can elicit this effect in machines as well as humans, so we have.

B

An English word for that. It's so jumpy when you react to things that you should necessarily react to yeah.

F

B

Yeah also means that you're more on guard yeah.

C

So the in like in more formal analysis of this, like when people study perception and they look at things in this framework like say they'll, take some signal like a tone or an image and they'll say you know, is it something above or below threshold like is it?

C

Is it elicit some feeling or not, and people have a range of different like sensitivities to it, so some people are much more sensitive to some stimulus than others and so they'll classified as a tone, even if most people don't- and so they can actually measure this on a curve as well in a performance curve. So even in people it's highly variable in terms of what they're responding to and so again yeah it's, it's not necessarily a false thing, but it's like you know.

C

It's just the way we recognize patterns so, but we can also you know these. These kind of things can also be elicited by malicious actors, and so you know you can creates a stimuli in a in a machine learning context that, where you produce something that looks like maybe a face, you know, maybe you want to replicate fingerprints or you want to replicate a face for some reason to fool what they call again. You know when they use these terms.

C

They do sort of conduct a certain things but they're just using them as a stand-in for different things that they want to convey. So why deep learning a eyes are so easy to fool is one example of an article where they show this sort of thing. They call it. You know they think of it in terms of model, weirdo, freddo, illness versus model robustness, but you could think of in terms of model. You know some sort of pareidolia response, depending on what you want to do. The.

B

Plant world does this. So if you take your picture plants which attract insects,.

B

C

But these are a bunch of articles that talk about this topic, so this is the interaction of this sort of pareidolia effect and people actually attacking systems like deep learning models and AIS for different things. This houdini flowing deep structure of prediction models is kind of interesting when this advant real-world adversarial attack on our face. So there's a thing in computer science called adversarial attack, which is where you use some sort of you know use.

C

Some sort of you know thought whether you make little false positive data set to sort of flood the zone in the yogur of them with false examples. So sometimes you can use adversarial techniques to improve the models such as any, but in this case they're using it to sort of elicit false results in a face. Id system, which is, if you you know, want to recognize faces correctly. Every time is a bad thing, but if you want to create a system that produces faces for art or something like that, it's actually a good thing.

C

So you know you can use these things in different ways, but this phenomenon can be elicited in machines and those are some examples of cases for that. So so here's another example. This is actually from this nature. Article deep learning a eyes are so easy to fool. This is an image from that particle and they talk about actually using DNS and being able to rotate objects to confuse DNS and again DM NS are really you know, dependent on this sort of false positive to positive system of classification.

C

So if you have this, you know you can actually rotate images like we saw in the inside tube or we looked at the the duck and the rabbit illusion. In this case you can rotate objects and you can elicit different responses from the system. So here we have a stop sign, and this is generally the orientation that we see stop signs that when we drive down the street or when a self-driving car drives down the street. Probably more importantly, for this example, you have the stop signs right-side up, it's all tilted. Maybe it turns out.

C

If you tilt it sideways, you get something called a it recognizes it as a dumbbell, which is you know not what it is, but that's all it's right and then, if you set up to excite and squish it, you know you do this rotation of the image and you you know skew the letters it actually recognizes as a racket.

C

So that's recognizing all these things out of one image, if you just change the orientation of the shape- and this of course is also you know- you can see like from the Mona Lisa example where you know, if you have these shapes that are kind of you know, shape you know squished or combined in different ways. You can elicit certain responses, and then you know in the nervous system, but also in a neural network.

C

Again, then we have these other examples. So you have. This is a an insect, a moth and it's sitting on a table cover.

C

So it's sitting on this surface and it's I, don't know how it's detecting a manhole cover, maybe from the table back underneath this mad, but it's recognizing it as such, and then this is my favorite, a mushroom, that's kind of twisted in the of its counting sideways and that's classified as a pretzel. It's because it's twisted like this, so you get this sort of you get this sort of effect where, if you change the orientation of these things, you know you change the identity of it. The Oh at least what the algorithm uses the identity.

C

So the the caption here is even natural images can fool a DNN because it might focus on the pictures. Color texture background, rather than picking out the salient features. So the salient features actually the way we process objects is a little bit different as well, then then faces you're picking out certain features in the image like corners or letters in the relationship between those, and so here we can actually see how that could be manipulated, or you know, change yeah.

B

That's it maybe what's missing from the da nose is: what's called mental rotation, there's a whole body of research going back Nick, it's a mental rotation, for example. If you took the stop sign, we relate example the amount of time it takes you to recognize that as long as the stop sign is longer, if you have to rotate it back to the standard orientation in your reading- and you can actually measure this so I'm going to suggest that difference here is we recognize this, but first we rotate it to a canonical position.

B

Okay, maybe that's what's lacking.

B

C

Mate, well, that's interested in interest at a point because yeah they don't do that on their own they're. There ways that you can train these models so that they.

B

It's not the same as testing if you think you see versus if you take one way that people test this idea, is you take an image of something that whose mirror image is not the same as the object like I'd like two hands, and then you present this at an on both angle.

B

Then you ask: is it the same or is it the mirror image and the things to the person to make the decision depends on the ambulance for notated, and so there's actually seems to be a rotation algorithm going on in your brain that iterates until you see cific, to make a match or not, and that's not the same as be presented with the object at many different orientations. So you can just select that it's there yeah.

C

B

So I'm suggesting is that maybe what the ends need is the addition of mental rotation, software yeah.

C

Yeah, you know cuz in a DNN or maybe like in any neural network. You really have to Train it in that way. So you know you might get a good set of results if the images are one orientation, but if you, if they're in different contexts, different backgrounds, even they'll, have the same problem. Yeah.

B

Is what I'm saying is you mental rotation? You know what you know. We recognize the object enough to know that it has a canonical representation and then you can actually add is the. Is that image visit canonical or not yeah.

C

B

If you just have, if I just have a hundred images of a hundred different angles and I just have to select, which one that's very different from the mental rotation.

B

Mental rotation in vienna's I.

C

I'm not run across I, didn't really search for that specific thing, but I don't know of any examples. This.

G

C

Only thing I know is that they, you know they do have problems with like generalizing data, so it really is like from what I know is that people, unless they train them specifically for that, but no one has come up with an algorithm that has like a separate module for mental rotation.

C

So those are all the slides I have for that. Do we have any questions in the let's see we have some things in the chat here so okay, Vinay says there has been some research on this topic. Can deep networks learn invariance?

C

So that's the link in the chat, and so they call it learning invariants how it should be. Ag, I, actually they're talking about here, how it should be able to learn, recognize new classes of objects from as few examples as possible.

C

So I mean, if you want to if people want to go through this and read it, it looks pretty interesting. So they've done some experiments with this okay, so we can check yeah, it's yeah. It should be right there. Let me make a link.

G

C

So they have something called spatial Transformers.

C

Current models capable learning spatial transformations to make recognition models, invariant.

C

So they have to know the class of transformations in advance for this, but then that doesn't solve your problem if you've run into things that you don't know. So if you have a you know a set of transformations like rotations around a circle, you know what those are, but if you squish it and stretch it out, that might not be something it'll work nice. So you have to train in advance of all those things as a separate module, and then you know this is always a tricky problem, trying to find things that general okay.

C

So that's all I have for that presentation, so Jessie missed the first part of it. But if you have any questions about it any thoughts, yes, the rotation models prefer more poorly. Is the rotation three increases.

E

I do learned a lot.

C

E

Topics here, I really I know them, but they not know that. There's a research topic behind these kinds of ambiguities, so I.

C

Mean it's not really so much a research topic as like just the thing that we can observe in nature, and then we also observe in algorithms I mean like I, gave you a paucity of preferences for it, so I, don't really think it's built. That was like a couple years ago, when I did that blog post and they don't think since then I was looking it up. I was, as I was doing. This talk and I didn't see much more than that out there. So it's definitely an open area. These are the slides here.

C

Definitely an open area. People were interested in thinking about it more, you know, I, don't think people have talked about it too much. That's a thing about like a lot of machine learning. Research is that there are a lot of like topics that kind of come from human cognition that you know kind of apply to these models, because we're we're not recreating a nervous system necessarily but we're creating the conditions for cognition, and so there are things in cognition that are really interesting.

C

That pop up in these models that no one's really thought about too much so.

C

So I'm Jessie did you have any questions or comments.

B

C

Yeah, well, it's like the term pareidolia. They use it in cognition, but then, when they talk about it in computers, it's slightly different like so. If you find someone who studied visual perception, they might say. Well, that's not pareidolia, that's something else, but you know point taken, but they don't know have a better word for it. So.

C

Why don't we switch gears then, and talk about the so we're we're nick and I were working on the digital battle area paper? And so what are your thoughts on that I saw? You send an email about splitting they're having a series of two papers? Oh.

B

The second would be on psychoanalysis.

B

I think what we're headed for here is that the single colony is actually an organist yeah.

C

B

Which fits with the original description about a turn on the beach.

B

Okay, so I think all the business about trying to figure out what algorithm is the individual self doing, so that it can cooperate with all the others to produce these different which pull dynamic, phenotypes yeah and that's what I see is the second paper.

C

Yeah I think that's a good idea. Okay,.

F

C

But like in this in this in this paper, you do you think it needs more of anything like specifically that we are missing.

B

Well, the only thing that I saw missing is that for the person who does not is not from the least current state of the art, an image processing be nice to take one image, and you talked about a number of steps. The goes through show the result, all the intermediates stages.

B

Okay, I, think! That's! That's the only thing. That's all that's missing, so that one can see how the algorithm is through the various ends up producing result where, after at least in most cases and also show word fast now, you pointed out that it's the so if the cells are all strung out so they're almost in align, you kind of lose it yeah.

C

B

Other thing that's missing is I, didn't quite understand when you kept using the word static. Oh.

C

A

It out, but what did you mean Oh.

C

Individual images are static: oh okay,.

B

In essence, yeah, that's a single frame from a movie; it's not static. It's just a single time. Point yeah.

F

B

It's not static, the static implies it doesn't change over time. Okay, okay, this way, I guess that's what it's a single frame and movie, but what is not used as the information, for example, that the number of cells is the same from one frame to the next and therefore one could say: if number of frames has changed, one of two things has happened: the algorithm screwed up or there's been a cell division right.

B

Okay, so so a higher order algorithm would take into account consecutive frames and the number of cells that yet okay can I correct a frame based on the frame before the framing after it. Unless I think I protected a subdivision great now, a television I think could be detectable by the cells. Getting whiter ya know those the the width rather than length would increased, so I think.

B

Maybe you could actually objectively determine that a cell is going through the cell division process and that should actually take many a long time compared to the time it takes for the whole colony to move back and forth. Yeah.

B

Okay, so is the amount of time it takes. Is a per cell division that was also a criterion for being a cell division.

B

So I guess that would require plotting the width of a cell versus time and anything to see if it's changing for some of the cells and not for others.

C

F

B

Ask fun questions like: are the cell divisions, Sigma innocence.

B

Okay, so the higher the higher orders but right I think a little bit of discussion about how one might incorporate the time-dependent. What were these were signing or time independent parameters that could be used to correct a single frame, yeah.

C

F

B

Mean there's probably a lot of work by itself, but just you know future work, sections yeah.

C

C

Yeah I thought the end probably needed some more detail, but I didn't know exactly how to go about doing that. But I think I know now: okay,.

B

I used to have a strong interest in mental rotation that list of 40 papers are so here. Okay, anybody interested say sure.

B

Okay, nice anything a file here or no, no.

C

But you can exchange it by email, yeah.

B

B

It's not okay, I'll tell you what Dennis takes the time to bring it up to date, see.

C

G

Looked at this, it's 2014.

B

References on gender differences, mental rotation processes.

C

So yeah I mean that's good. Thank you for the update on the paper and so what's the deadline for submission.

B

We gotta leave some time for review. I, say push. It would be nice to submit it within a couple of weeks. Okay, okay and certainly I- think we're pretty close, so I'd like to submit by October 28th that possible, because we've got a meeting coming up on that date with the editors to his best things.

B

We can make that understand this.

G

C

Jesse said yeah about the two papers, so these are the that we were referring to as the digital basilar area papers.

C

This is from the other group that was working with this jessie was in the friedenberg vehicles group, but we also had a precursor to this group, which was the google summer code machine learning group, and then we had the one group working in digital vessel area and that's the paper that this is based on so I presented on that, like in week, 2 of the Ivoire ml group Jesse, where we have these LG that are these diatoms that have a bunch of filaments that so I had to.

C

You know slide together in a colony and we're analyzing images of those organisms so and then dick proposed. You know having like a second paper to follow up on this one, but that's something that would be later on, and so you know we're you know, that's what you get when you do research, you get more research from the research that you've already done so yeah. Oh.

B

By the way, Thomas Harbhajan look at yet ah no.

C

I'll, send it to him. Okay,.

B

Thanks thanks getting the point where he could go over yeah.

C

Yeah because I think I have to point it out to him as pacifically yeah I. Don't think he's seen it, though.

C

So I wanted to talk in the last part of the meeting about the pre-trained model. Was blog post, so I actually worked on flushing this out, I'm gonna present the screen. So actually let me put the link to the document in the chat, so we have a common reference point here, and this is the blogpost I talked about. This is actually from the pre-training model. Discussion we've had in the group and the day and I were talking about this yesterday.

C

So this last week, I was able to flesh out a lot of the details that I last week, I presented an outline, and this week I've got a draft here, and so the outline was based on. You know three different sets of topics and we're so we're heading towards a 15 under word limit. So right now, I think we're at 750.

F

C

Words so we're well within our limit, so we have room for people to contribute things if they want. So this starts out with an introduction to pre train models, and this could use some work. I just gave a couple of examples here, and we talked a little bit about you know what the advantages are for biologists and then a vision for sort of a developmental biology, specific model, and so we kind of contemplate what those what the advantages of such a pre training model might be.

C

So there are a couple of features of developmental systems that are bullet pointed here that are relevant, I mean maybe I got these off the top of my head. So if you can rephrase these or if you want to change them or add to them, that would work as well, but I was basing these points on the just general things in developmental biology that are unique to that type of system.

C

Maybe not unique but like definitely needs need to be addressed like cell division and differentiation events, the ability to identify a cell and when it you know, divides when it replicates, but also as its as its replicating. So to be able to identify thing features like that. um Then we have some discussion about models.

C

Not equaling phenomenology, which is you know, has implications for the model and I kind of wrote that out a bit and then this is the part that I'm not done with you and I'm gonna work on later kind of giving some practical examples, and that's like it begin as I told the neh. This is a place where we can bring a lot of this stuff to bear from say, like the the google Summer of Code, and also some of you know, practical examples from actual data sets.

C

So you know we might create a hypothetical set of biological images in our head, or we might like think about like a process or something I'm, not quite sure what the what it'll look like. But um you know I know. Vinay said he was gonna work on it today in tomorrow, and then Jesse Indic. If you want to contribute to this you're welcome, but you're not required to do so. All right, yes, even going over it and making corrections would be helpful too, because we're trying to get this published.

C

We're gonna publish this on the note, which is the developmental biology blog. They need sometimes this or next week. So you want to kind of wrap this up fairly quickly. So if we can get some feedback, that would be great or some contributions. That would be great as well yeah just he thinks yeah definitely take a look at it and even if you just want to do some copy editing, that's that's welcome and they you know appreciate your input, so I mean you know.

C

I think we again like this is based on the outline that we had Oh.

C

Dick fell out of the meeting.

B

Something-Something drones, my rotor, oh okay,.

C

Sorry I, just I was just talking about this preacher model blogpost and if you know, if you have any, if you want to look it over and have any suggestions- or you know, copy editing suggestions or whatever that's welcome, we're gonna try to put it out on this blog, the node, which is a developmental biology blog, but maybe like the beginning of next week. So and then we might do another blog post later on year, I'm not sure yet we'll see how you know how this one goes over.

C

I hope we can get some people to read it and get interested in it. This will be definitely one of the you know. I mentioned the group on the node and I. Don't know how many people saw the post but like to keep them updated a little bit on what we've been talking about so so yeah. That would be good. If people want to go over that again, it took miss the link. This is the link.

C

Pavan a did. You have any questions you you. Actually. It asked me specifically about this.

E

I want to get an idea of how the content should look like like one of the types we can add or the extra content that we can actual the region, it's in view as well so, first time I'm reading the entire document, then I will enter flow of how to continue this and then probably I had when I was doing.

E

This I came across some printed models, for especially for the medical field, let's say for brain scanning a this part of any segmentation kind of especially for the health candy from the healthcare field, so I like to use that and then I had some introduction to those models and then the next to them, yeah yeah. So.

C

Yeah one more point about that: we have like the.

F

Switches so.

C

What I've learned from doing blog posts is you want to keep the descriptions brief in the main part of the blog post like if you do something you know you can make a brief description of something, that's good, but for more detailed things, we have like an appendix area, which is this resources section. It's actually I thumb this one, the model zoo, which is a nice little reference on different pre train models for a wide variety of things, but we can also put like resources for like different specific models and kind of annotating.

C

You know so we like mentioned the model, mention what it's used for and then provide a link and then have like a list of those down here. That way, people can like go through the blog post or I quickly.

C

Read it digest it and then they go down and they can find some references and some resources and they don't have to like, be they don't have to have in the way when they're reading it I find blogs that people are very impatient when they read blog posts, they want to get cut to the chase and then, like you know, you'll have the resources at the end, so I just found that to be a good strategy. For that so I mean you know, keep that in mind when you're going through these examples.

C

You know if it's something that maybe you want to put in the resources section or actually, like you, explain it in the text. If you can do a very maybe in a couple sentences summarize it then that's fine, but otherwise we might put it in the latter section.

E

C

Short things: yeah, okay, I did anyone else, have any questions about it or Jessie? Did you have any thoughts about it?.

C

Okay, yeah I mean yeah, take a look, and then you know slack me or email me about it and yeah. That's good! So yeah, that's good! Well, that's I! Think that's our meeting for today. Unless anyone has any other things they want to mention at the top of the hour.

F

C

Well: okay! Well, thanks for attending next week is open. We don't have anything yeah thanks Jesse, we don't have anything on the schedule. So if you want to.