gberseth/IFT6758-comments · Datasets at Hugging Face

input stringlengths 0 513	output stringlengths 0 513
start	I might have to tell all of you to drop all of your Wi-fi signals in here. Just so it doesn't crash this tablet.
I might have to tell all of you to drop all of your Wi-fi signals in here. Just so it doesn't crash this tablet.	But anyway.
But anyway.	should be good for now, just for a few more
should be good for now, just for a few more	gaps
gaps	here. Okay?
here. Okay?	And you can use these types of models to generate like synthetic images which we've talked about a bit from before
And you can use these types of models to generate like synthetic images which we've talked about a bit from before	they can actually generate some pretty nice ones. These days. People are using diffusion models more than Gans, because the diffusion models are just way, less of a headache to be able to train, and they tend to actually map these distributions much nicer.
they can actually generate some pretty nice ones. These days. People are using diffusion models more than Gans, because the diffusion models are just way, less of a headache to be able to train, and they tend to actually map these distributions much nicer.	So kind of like. How I was saying, if we train a generative model, and I used to pick up
So kind of like. How I was saying, if we train a generative model, and I used to pick up	on like the average data in the middle, a lot more Gans are pretty guilty of that problem. For sure, the fusion models are a little bit better.
on like the average data in the middle, a lot more Gans are pretty guilty of that problem. For sure, the fusion models are a little bit better.	and this is where we can at least take
and this is where we can at least take	now these models that are generating images. Try to plug in a little piece to it that says, Please now generate something in the style of this image.
now these models that are generating images. Try to plug in a little piece to it that says, Please now generate something in the style of this image.	and that'll work pretty well.
and that'll work pretty well.	and you can do this for a whole bunch of different styles. So instead of just doing
and you can do this for a whole bunch of different styles. So instead of just doing	translation, you can try and style it such that it looks like it's in the night. Maybe it's a picture in the winter, or something like that.
translation, you can try and style it such that it looks like it's in the night. Maybe it's a picture in the winter, or something like that.	So you can generate a lot more data.
So you can generate a lot more data.	Yeah, like a summer winter switch.
Yeah, like a summer winter switch.	And then there's a recent piece of work which hopefully, I'll be able to get to.
And then there's a recent piece of work which hopefully, I'll be able to get to.	Here we go
Here we go	From a previous student of mine where you can use now like text to image diffusion models.
From a previous student of mine where you can use now like text to image diffusion models.	So if it's got a diffusion model for generating images.
So if it's got a diffusion model for generating images.	now, you can feed in some text such that you can identify. Okay, this is of class
now, you can feed in some text such that you can identify. Okay, this is of class	train. Now please generate me a bunch of images that are essentially of the same class.
train. Now please generate me a bunch of images that are essentially of the same class.	because maybe you add a bunch of stuff in the data set
because maybe you add a bunch of stuff in the data set	of a bunch of other metallic things that travel on wheels. In some way, you can figure how to generate some stuff from this distribution.
of a bunch of other metallic things that travel on wheels. In some way, you can figure how to generate some stuff from this distribution.	So essentially, this is figuring out how to key in. All right. If you've got a particular class, it's trying to share information across classes such that it can generate some more data in that particular class. So interesting type of data augmentation. You can go have a look at this
So essentially, this is figuring out how to key in. All right. If you've got a particular class, it's trying to share information across classes such that it can generate some more data in that particular class. So interesting type of data augmentation. You can go have a look at this	paper. If you want to know more about this particular model.
paper. If you want to know more about this particular model.	it is a bit tricky to actually train.
it is a bit tricky to actually train.	He shares his final chicken version
He shares his final chicken version	like.
like.	no, okay.
no, okay.	I'll send a message.
I'll send a message.	and I'll update the version of this right after. Yeah, I might have added this on the weekend
and I'll update the version of this right after. Yeah, I might have added this on the weekend	and forgot to upload this one.
and forgot to upload this one.	Otherwise you can just have a look for effective data augmentation with the fusion models
Otherwise you can just have a look for effective data augmentation with the fusion models	from Brandon Chabuco.
from Brandon Chabuco.	Otherwise, if I don't post the updated slides by tonight. Someone message me on fiance
Otherwise, if I don't post the updated slides by tonight. Someone message me on fiance	and I'll make sure to put them up.
and I'll make sure to put them up.	Okay?
Okay?	So those were a lot of methods for being able to do image data augmentation.
So those were a lot of methods for being able to do image data augmentation.	Like, I said, there's always a few challenges for being able to do this for tabular data. It's just not as continuous. So our data distributions are a bit more tricky to work with.
Like, I said, there's always a few challenges for being able to do this for tabular data. It's just not as continuous. So our data distributions are a bit more tricky to work with.	So if each point is basically, if each entry in a table.
So if each point is basically, if each entry in a table.	it's just in our like d dimensional space. Now.
it's just in our like d dimensional space. Now.	and we want to generate some new points while respecting some boundaries. Indeed!
and we want to generate some new points while respecting some boundaries. Indeed!	Then we can figure out how to generate some more data.
Then we can figure out how to generate some more data.	Often this ends up coming out because we know something about the data set. So you can maybe apply Gaussian noise to
Often this ends up coming out because we know something about the data set. So you can maybe apply Gaussian noise to	generate some more data, but also add some Gaussy noise to the age. But we should know that no one should have the age of Niagara , and no one should be born in.
generate some more data, but also add some Gaussy noise to the age. But we should know that no one should have the age of Niagara , and no one should be born in.	or something like that. So usually we can apply the filters
or something like that. So usually we can apply the filters	on top of the very generic data, and then we can use that rejection sampling method again. So if you accidentally produce one like this, okay, garbage. Generate another one and then keep that part in the data. Set
on top of the very generic data, and then we can use that rejection sampling method again. So if you accidentally produce one like this, okay, garbage. Generate another one and then keep that part in the data. Set	one more time.
one more time.	Okay?
Okay?	And this is where you can use kind of like the classic smoke method does a bit of a version of this sampling in order to get some points there.
And this is where you can use kind of like the classic smoke method does a bit of a version of this sampling in order to get some points there.	And there's a few more
And there's a few more	advanced method. If you're getting into this kind of uglier area of having to deal with data sets and lack of data
advanced method. If you're getting into this kind of uglier area of having to deal with data sets and lack of data	you can use.
you can use.	I think this is pronounced at a sign. But I'm not really sure, and this will even randomly modify some of the positions of even these generated points.
I think this is pronounced at a sign. But I'm not really sure, and this will even randomly modify some of the positions of even these generated points.	And there's a few examples of using tabular gans. And actually, there's a couple of good new, like really recent pieces of the research week. Because these things are often text actually using Llms is a good application for figuring how to generate some more tabular data.
And there's a few examples of using tabular gans. And actually, there's a couple of good new, like really recent pieces of the research week. Because these things are often text actually using Llms is a good application for figuring how to generate some more tabular data.	It's of the few things that works well for text and tabular data.
It's of the few things that works well for text and tabular data.	Those Lms like typically deep learning, has not been a good thing to apply on tabular data.
Those Lms like typically deep learning, has not been a good thing to apply on tabular data.	because tabulate data is either too discrete or there's not enough of it.
because tabulate data is either too discrete or there's not enough of it.	And that's where still often, like somewhat more traditional methods, work the best there.
And that's where still often, like somewhat more traditional methods, work the best there.	Okay.
Okay.	Golden Rule.
Golden Rule.	If you can convince your boss or yourself or your funders to give you some money for more data, that's usually the best way to do things. So go with that if you can.
If you can convince your boss or yourself or your funders to give you some money for more data, that's usually the best way to do things. So go with that if you can.	And then the last thing yeah, which I'll cover in just a minute
And then the last thing yeah, which I'll cover in just a minute	is data drift. And this is where you kind of like, it seems like you have multiple classes of data.
is data drift. And this is where you kind of like, it seems like you have multiple classes of data.	But you actually just are doing something over some sort of data drives.
But you actually just are doing something over some sort of data drives.	So let's see if you've got this kind of compensation common situation.
So let's see if you've got this kind of compensation common situation.	your company, you're gonna collect data over the last months.
your company, you're gonna collect data over the last months.	And then what you're gonna end up deciding to do. You're gonna take the st months of that data as training and validation in the last month as test for some reason. Remember, don't do that. We wanna kind of split things equally so. Everything looks like the same distribution.
And then what you're gonna end up deciding to do. You're gonna take the st months of that data as training and validation in the last month as test for some reason. Remember, don't do that. We wanna kind of split things equally so. Everything looks like the same distribution.	But it's gonna perform really well on these sets, but very poorly on the test set.
But it's gonna perform really well on these sets, but very poorly on the test set.	So we want to figure out what's actually going on here.
So we want to figure out what's actually going on here.	The problem is okay. These weren't part of it didn't end up being part of the same distribution, not necessarily because of anything you did, but
The problem is okay. These weren't part of it didn't end up being part of the same distribution, not necessarily because of anything you did, but	because we all currently operate on the well, the idea that we exist in linear time.
because we all currently operate on the well, the idea that we exist in linear time.	and that there are other things that will happen because of that.
and that there are other things that will happen because of that.	So we can figure out, how can we detect drift and what to do about it. There's a lot of examples for this, you know.
So we can figure out, how can we detect drift and what to do about it. There's a lot of examples for this, you know.	or whatever. Years ago we'd all be teaching remotely right now, and there's a whole bunch of weird this print distribution during Covid. So if you collected your data during that time, I don't know. Maybe you should throw it out now, and we're all back to normal. Supposedly.
or whatever. Years ago we'd all be teaching remotely right now, and there's a whole bunch of weird this print distribution during Covid. So if you collected your data during that time, I don't know. Maybe you should throw it out now, and we're all back to normal. Supposedly.	And this will be some indication that our models underperforming.
And this will be some indication that our models underperforming.	So there is one nice way to be able to do this.
So there is one nice way to be able to do this.	To do some of these tests is you can do some antagonist validation.
To do some of these tests is you can do some antagonist validation.	So if you've got your test sets and your training set here, which are colored.
So if you've got your test sets and your training set here, which are colored.	You can actually like train
You can actually like train	if a classifier is able to differentiate between the sets. So before we talked about like you compute the mean over something. But now, if I train just a classifier label
if a classifier is able to differentiate between the sets. So before we talked about like you compute the mean over something. But now, if I train just a classifier label	the stuff from the things in the test set as zeros and the stuff that's in the train set as ones.
the stuff from the things in the test set as zeros and the stuff that's in the train set as ones.	It should not be able to get this right like it should end up being like wrong. or right, % of the time, because it just can't tell the difference.
It should not be able to get this right like it should end up being like wrong. or right, % of the time, because it just can't tell the difference.	That's way to do that.
That's way to do that.	This is more helpful, too. If your data is like not normally distributed. And you still want to figure out if it's in the same distribution.
This is more helpful, too. If your data is like not normally distributed. And you still want to figure out if it's in the same distribution.	And this is where, yeah, we'll have no drift if we basically can't figure out how to separate this data or some drift if it does exist.
And this is where, yeah, we'll have no drift if we basically can't figure out how to separate this data or some drift if it does exist.	And this happens a lot like according to weather or whatever. Or there's a lot of periodic functions that happen according to the week.
And this happens a lot like according to weather or whatever. Or there's a lot of periodic functions that happen according to the week.	Everyone does a lot of stuff or a lot different stuff on the weekend. It's also just a part of the way our society is structured.
Everyone does a lot of stuff or a lot different stuff on the weekend. It's also just a part of the way our society is structured.	So a lot of kind of odd things happen.
So a lot of kind of odd things happen.	yeah. And then
yeah. And then	you end up constructing this data set in some way. Remember, you're not going to look quite at the target. You're gonna look at this test set versus training. Set this distribution.
you end up constructing this data set in some way. Remember, you're not going to look quite at the target. You're gonna look at this test set versus training. Set this distribution.	And then the results must be compared to some sort of random mixture with some null hypothesis. A classifier will be, you know, worse at differentiating a random mixture that might also be a good check, too.
And then the results must be compared to some sort of random mixture with some null hypothesis. A classifier will be, you know, worse at differentiating a random mixture that might also be a good check, too.	So like
So like	this is the nice case we've got like zeros and ones, and different deficits.
this is the nice case we've got like zeros and ones, and different deficits.	You might also train basically some model that's trying to predict. Just like random stuff. You take all of your labels and just lay random, label them all randomly.
You might also train basically some model that's trying to predict. Just like random stuff. You take all of your labels and just lay random, label them all randomly.	just to check, as it's kind of a slightly better baseline than just taking %. Maybe your model can somehow guess on a few things for some class, or whatever
just to check, as it's kind of a slightly better baseline than just taking %. Maybe your model can somehow guess on a few things for some class, or whatever	it's a better. It's an additional way to be able to check this version of looking at different distributions.
it's a better. It's an additional way to be able to check this version of looking at different distributions.	And hopefully, if you're lucky you'll be in some space that's like a constant drift. No sorry.
And hopefully, if you're lucky you'll be in some space that's like a constant drift. No sorry.	a constant drift problem, which is also kind of hard. This is sort of

start

I might have to tell all of you to drop all of your Wi-fi signals in here. Just so it doesn't crash this tablet.

But anyway.

should be good for now, just for a few more

gaps

here. Okay?

And you can use these types of models to generate like synthetic images which we've talked about a bit from before

they can actually generate some pretty nice ones. These days. People are using diffusion models more than Gans, because the diffusion models are just way, less of a headache to be able to train, and they tend to actually map these distributions much nicer.

So kind of like. How I was saying, if we train a generative model, and I used to pick up

on like the average data in the middle, a lot more Gans are pretty guilty of that problem. For sure, the fusion models are a little bit better.

and this is where we can at least take

now these models that are generating images. Try to plug in a little piece to it that says, Please now generate something in the style of this image.

and that'll work pretty well.

and you can do this for a whole bunch of different styles. So instead of just doing

translation, you can try and style it such that it looks like it's in the night. Maybe it's a picture in the winter, or something like that.

So you can generate a lot more data.

Yeah, like a summer winter switch.

And then there's a recent piece of work which hopefully, I'll be able to get to.

Here we go

From a previous student of mine where you can use now like text to image diffusion models.

So if it's got a diffusion model for generating images.

now, you can feed in some text such that you can identify. Okay, this is of class

train. Now please generate me a bunch of images that are essentially of the same class.

because maybe you add a bunch of stuff in the data set

of a bunch of other metallic things that travel on wheels. In some way, you can figure how to generate some stuff from this distribution.

So essentially, this is figuring out how to key in. All right. If you've got a particular class, it's trying to share information across classes such that it can generate some more data in that particular class. So interesting type of data augmentation. You can go have a look at this

paper. If you want to know more about this particular model.

it is a bit tricky to actually train.

He shares his final chicken version

like.

no, okay.

I'll send a message.

and I'll update the version of this right after. Yeah, I might have added this on the weekend

and forgot to upload this one.

Otherwise you can just have a look for effective data augmentation with the fusion models

from Brandon Chabuco.

Otherwise, if I don't post the updated slides by tonight. Someone message me on fiance

and I'll make sure to put them up.

Okay?

So those were a lot of methods for being able to do image data augmentation.

Like, I said, there's always a few challenges for being able to do this for tabular data. It's just not as continuous. So our data distributions are a bit more tricky to work with.

So if each point is basically, if each entry in a table.

it's just in our like d dimensional space. Now.

and we want to generate some new points while respecting some boundaries. Indeed!

Then we can figure out how to generate some more data.

Often this ends up coming out because we know something about the data set. So you can maybe apply Gaussian noise to

generate some more data, but also add some Gaussy noise to the age. But we should know that no one should have the age of Niagara , and no one should be born in.

or something like that. So usually we can apply the filters

on top of the very generic data, and then we can use that rejection sampling method again. So if you accidentally produce one like this, okay, garbage. Generate another one and then keep that part in the data. Set

one more time.

Okay?

And this is where you can use kind of like the classic smoke method does a bit of a version of this sampling in order to get some points there.

And there's a few more

advanced method. If you're getting into this kind of uglier area of having to deal with data sets and lack of data

you can use.

I think this is pronounced at a sign. But I'm not really sure, and this will even randomly modify some of the positions of even these generated points.

And there's a few examples of using tabular gans. And actually, there's a couple of good new, like really recent pieces of the research week. Because these things are often text actually using Llms is a good application for figuring how to generate some more tabular data.

It's of the few things that works well for text and tabular data.

Those Lms like typically deep learning, has not been a good thing to apply on tabular data.

because tabulate data is either too discrete or there's not enough of it.

And that's where still often, like somewhat more traditional methods, work the best there.

Okay.

Golden Rule.

If you can convince your boss or yourself or your funders to give you some money for more data, that's usually the best way to do things. So go with that if you can.

And then the last thing yeah, which I'll cover in just a minute

is data drift. And this is where you kind of like, it seems like you have multiple classes of data.

But you actually just are doing something over some sort of data drives.

So let's see if you've got this kind of compensation common situation.

your company, you're gonna collect data over the last months.

And then what you're gonna end up deciding to do. You're gonna take the st months of that data as training and validation in the last month as test for some reason. Remember, don't do that. We wanna kind of split things equally so. Everything looks like the same distribution.

But it's gonna perform really well on these sets, but very poorly on the test set.

So we want to figure out what's actually going on here.

The problem is okay. These weren't part of it didn't end up being part of the same distribution, not necessarily because of anything you did, but

because we all currently operate on the well, the idea that we exist in linear time.

and that there are other things that will happen because of that.

So we can figure out, how can we detect drift and what to do about it. There's a lot of examples for this, you know.

or whatever. Years ago we'd all be teaching remotely right now, and there's a whole bunch of weird this print distribution during Covid. So if you collected your data during that time, I don't know. Maybe you should throw it out now, and we're all back to normal. Supposedly.

And this will be some indication that our models underperforming.

So there is one nice way to be able to do this.

To do some of these tests is you can do some antagonist validation.

So if you've got your test sets and your training set here, which are colored.

You can actually like train

if a classifier is able to differentiate between the sets. So before we talked about like you compute the mean over something. But now, if I train just a classifier label

the stuff from the things in the test set as zeros and the stuff that's in the train set as ones.

It should not be able to get this right like it should end up being like wrong. or right, % of the time, because it just can't tell the difference.

That's way to do that.

This is more helpful, too. If your data is like not normally distributed. And you still want to figure out if it's in the same distribution.

And this is where, yeah, we'll have no drift if we basically can't figure out how to separate this data or some drift if it does exist.

And this happens a lot like according to weather or whatever. Or there's a lot of periodic functions that happen according to the week.

Everyone does a lot of stuff or a lot different stuff on the weekend. It's also just a part of the way our society is structured.

So a lot of kind of odd things happen.

yeah. And then

you end up constructing this data set in some way. Remember, you're not going to look quite at the target. You're gonna look at this test set versus training. Set this distribution.

And then the results must be compared to some sort of random mixture with some null hypothesis. A classifier will be, you know, worse at differentiating a random mixture that might also be a good check, too.

So like

this is the nice case we've got like zeros and ones, and different deficits.

You might also train basically some model that's trying to predict. Just like random stuff. You take all of your labels and just lay random, label them all randomly.

just to check, as it's kind of a slightly better baseline than just taking %. Maybe your model can somehow guess on a few things for some class, or whatever

it's a better. It's an additional way to be able to check this version of looking at different distributions.

And hopefully, if you're lucky you'll be in some space that's like a constant drift. No sorry.

a constant drift problem, which is also kind of hard. This is sort of