input stringlengths 0 513 | output stringlengths 0 513 |
|---|---|
model is, gonna st learn to assume. Okay, so the data is distributed this way. So I'll just predict blue. | Well, I don't know at least % of the time. |
Well, I don't know at least % of the time. | but the assumption is is that |
but the assumption is is that | it was kind of like something wrong with our data set. And if we had like a better sensor, or something like that, or more time, it would actually be even like existing in the real world. |
it was kind of like something wrong with our data set. And if we had like a better sensor, or something like that, or more time, it would actually be even like existing in the real world. | So this is. |
So this is. | I think, on the next slide. But I can say it now. It's just sort of like, you know, car crashes like, if you're collecting self data. You have, like |
I think, on the next slide. But I can say it now. It's just sort of like, you know, car crashes like, if you're collecting self data. You have, like | a hundred , times more data over here and be like |
a hundred , times more data over here and be like | smooth no issues column, and then barely anything in the car crash column, but they do happen, and we do care about them. |
smooth no issues column, and then barely anything in the car crash column, but they do happen, and we do care about them. | And it's not really that they don't exist. I'm sure if we all drove completely, randomly on the road. Maybe these distributions would be the uneasy, but because we don't do that, and we don't want to smash into each other all the time we end up with this very split difference? |
And it's not really that they don't exist. I'm sure if we all drove completely, randomly on the road. Maybe these distributions would be the uneasy, but because we don't do that, and we don't want to smash into each other all the time we end up with this very split difference? | Does that make sense? Any other questions about? |
Does that make sense? Any other questions about? | This is just kind of |
This is just kind of | something about the way we're collecting data leads us to have something that we know that the data set is a lot of data or a limited data |
something about the way we're collecting data leads us to have something that we know that the data set is a lot of data or a limited data | for the tax or for the picture. |
for the tax or for the picture. | Usually like, you'll know, like in the classification task, someone's gonna tell you to question based on these different classes. |
Usually like, you'll know, like in the classification task, someone's gonna tell you to question based on these different classes. | and then you should at least be able to compute over your data set. Okay, what is the percentage of data that I have that exists for each one of the classes? |
and then you should at least be able to compute over your data set. Okay, what is the percentage of data that I have that exists for each one of the classes? | And then |
And then | some of them are somehow not very even compared to the other classes, and you know that some invales |
some of them are somehow not very even compared to the other classes, and you know that some invales | sometimes we'll know a little bit already, depending on some of the extra knowledge we have as well. So that's why I use the software example |
sometimes we'll know a little bit already, depending on some of the extra knowledge we have as well. So that's why I use the software example | just because I know |
just because I know | I don't know. Cars would be useless if they split % of the time we use them, they all smash into each other so they must not. They must have some sort of weird. |
I don't know. Cars would be useless if they split % of the time we use them, they all smash into each other so they must not. They must have some sort of weird. | We can do |
We can do | good |
good | samples for the minority class, but which don't fit which we're not making it so they're very different in the example. And you just have to copy those. Will the model be able to figure out? Yes. |
samples for the minority class, but which don't fit which we're not making it so they're very different in the example. And you just have to copy those. Will the model be able to figure out? Yes. | but most often we need to apply like one more Band-aid over the problem |
but most often we need to apply like one more Band-aid over the problem | which I think is what's on the next slide? |
which I think is what's on the next slide? | No, we'll get to that in about slides. But basically, that's what we're going to do. Some sort of data augmentation, sometimes something as simple as just adding Gaussian noise. So they each point when you do a batch, look uniquely the same. But the underlying structure. |
No, we'll get to that in about slides. But basically, that's what we're going to do. Some sort of data augmentation, sometimes something as simple as just adding Gaussian noise. So they each point when you do a batch, look uniquely the same. But the underlying structure. | And we actually want them all to pick up a little bit. |
And we actually want them all to pick up a little bit. | Yeah. |
Yeah. | okay, so maybe I'll go ahead because that's basically what some of the next slides are for. |
okay, so maybe I'll go ahead because that's basically what some of the next slides are for. | So you can do the down sampling example, which is basically end up here on the right. Even if you have a lot more blue points. Let's just make them equal, and you can trade in your model. |
So you can do the down sampling example, which is basically end up here on the right. Even if you have a lot more blue points. Let's just make them equal, and you can trade in your model. | You can do this kind of interesting one where you can subsample fomec links, which is just alright. If you've got classes, and you find some points of them that are really similar to each other. |
You can do this kind of interesting one where you can subsample fomec links, which is just alright. If you've got classes, and you find some points of them that are really similar to each other. | you should probably sample those more. |
you should probably sample those more. | Just because that'll mean, like, I want my model to understand the difference between these the best. These other ones are like all the easy things. |
Just because that'll mean, like, I want my model to understand the difference between these the best. These other ones are like all the easy things. | Essentially, if it's hard to figure out the decision here, we can feed them all of those examples more. |
Essentially, if it's hard to figure out the decision here, we can feed them all of those examples more. | It's kind of I don't know. I mean, this is why we like design classes. I'm not talking random topics up here every day. I'm trying to specifically talk about the issues for data science. |
It's kind of I don't know. I mean, this is why we like design classes. I'm not talking random topics up here every day. I'm trying to specifically talk about the issues for data science. | There's this other version, which is. |
There's this other version, which is. | oh. |
oh. | so there's a version of this called Slopes. And it's really okay, if I want to do data augmentation. But I've got tabular data. So it's not as easy as that Gaussian noise to it. This you can look into the details for the smoke method to figure out, basically how to generate some points between these classes, such that... |
so there's a version of this called Slopes. And it's really okay, if I want to do data augmentation. But I've got tabular data. So it's not as easy as that Gaussian noise to it. This you can look into the details for the smoke method to figure out, basically how to generate some points between these classes, such that... | Just a couple of |
Just a couple of | did you say that, for example? |
did you say that, for example? | Yes. But I guess |
Yes. But I guess | this. Yeah, sorry about that. So in this case, if you've got your original data set, and you can identify |
this. Yeah, sorry about that. So in this case, if you've got your original data set, and you can identify | if you can identify these individual kind of parts that are close to each other. You can remove those from the asset |
if you can identify these individual kind of parts that are close to each other. You can remove those from the asset | as well. And then you can just speed up these different point. |
as well. And then you can just speed up these different point. | Yeah. |
Yeah. | okay, so this is a way of doing this. This often makes a very sharp model. But it's not going to do well on the points that are between these people look like |
okay, so this is a way of doing this. This often makes a very sharp model. But it's not going to do well on the points that are between these people look like | it's not a fantastic solution. |
it's not a fantastic solution. | The other the other way would be to |
The other the other way would be to | instead of like removing them. Do you just |
instead of like removing them. Do you just | generating more. |
generating more. | And with this phone specialist and then do that tomorrow, so you can go |
And with this phone specialist and then do that tomorrow, so you can go | here. |
here. | So |
So | so I would say that we haven't talked about it much, but a generative adversarial networks used to do something very similar to this. |
so I would say that we haven't talked about it much, but a generative adversarial networks used to do something very similar to this. | where you would create one model trying to differentiate between these . And then you create another machine model to try to generate something that confuses that model. |
where you would create one model trying to differentiate between these . And then you create another machine model to try to generate something that confuses that model. | We're probably along this boundary, a very tricky process for the administrators. |
We're probably along this boundary, a very tricky process for the administrators. | For that one way you could go about it. |
For that one way you could go about it. | But generally, though the reason why using this one is really hard, is it also assumed is that some sort of distance function in the state space already to identify which one of these are closer to other points. And if we have that distance function, we've basically already solved our machine learning problems. |
But generally, though the reason why using this one is really hard, is it also assumed is that some sort of distance function in the state space already to identify which one of these are closer to other points. And if we have that distance function, we've basically already solved our machine learning problems. | So it's it's just easier when we've got some. Someone's giving us a different function already. Or it's very |
So it's it's just easier when we've got some. Someone's giving us a different function already. Or it's very | obvious. |
obvious. | And look at clients like in this image and see that they're missing. |
And look at clients like in this image and see that they're missing. | or at least similar to each other. |
or at least similar to each other. | It's kind of like if you did, Pca. And found stuff that was close to each other. Well, Pca already did % of your work. |
It's kind of like if you did, Pca. And found stuff that was close to each other. Well, Pca already did % of your work. | Yeah, sorry earlier. |
Yeah, sorry earlier. | said, if you have data this quarter, I'll share the latest. |
said, if you have data this quarter, I'll share the latest. | It's why is it? |
It's why is it? | I would think that kind of data. |
I would think that kind of data. | Well, what if one of your data types in your tabular data are the colors like colors |
Well, what if one of your data types in your tabular data are the colors like colors | or something like that. So it's not easy to just add Gaussian noise to something that is essentially a categorical variable. |
or something like that. So it's not easy to just add Gaussian noise to something that is essentially a categorical variable. | So it's the challenge of just doing stuff of categorical values. |
So it's the challenge of just doing stuff of categorical values. | And that's what makes this sort of subsampling process a bit harder. |
And that's what makes this sort of subsampling process a bit harder. | She can use that |
She can use that | So the nice part well again, if you're lucky, you can do a nice data increase instead. |
So the nice part well again, if you're lucky, you can do a nice data increase instead. | So we mentioned, you can do this over sampling of the classes that basically don't have nearly as much data. |
So we mentioned, you can do this over sampling of the classes that basically don't have nearly as much data. | and there's a lot of examples of these that we sort of talked about one example already. |
and there's a lot of examples of these that we sort of talked about one example already. | But |
But | by examples, I hear. I mean, when we have really unbalanced data sets or like classes in our data, there's a lot of finance problems for detecting fraudulent transactions. |
by examples, I hear. I mean, when we have really unbalanced data sets or like classes in our data, there's a lot of finance problems for detecting fraudulent transactions. | At least, according to myself, there's a lot of marketing stuff where you only click on stuff. Very rarely. |
At least, according to myself, there's a lot of marketing stuff where you only click on stuff. Very rarely. | I basically avoid clicking on any, ever, so that always probably makes my feedback really confusing, which is how I like it. |
I basically avoid clicking on any, ever, so that always probably makes my feedback really confusing, which is how I like it. | I would hope that medical stuff will be very rare as well. So any kind of cancer detection is also hard. |
I would hope that medical stuff will be very rare as well. So any kind of cancer detection is also hard. | And oh. |
And oh. | spam emails used to be something that were very rare. But now, like maybe % of your emails, you get or spam emails |
spam emails used to be something that were very rare. But now, like maybe % of your emails, you get or spam emails | so they can be kind of. |
so they can be kind of. | Now, my like real to spam ratio is probably inverted than when it was when I was younger. |
Now, my like real to spam ratio is probably inverted than when it was when I was younger. | So whatever |
So whatever | there's also some examples for image data I gave that Lamborghini example or industry security. |
there's also some examples for image data I gave that Lamborghini example or industry security. | So does that make sense like these do exist a lot. And it's usually where machine learning becomes really helpful. |
So does that make sense like these do exist a lot. And it's usually where machine learning becomes really helpful. | and it's often where we want to use outlier detection to actually find stuff we're interested in. |
and it's often where we want to use outlier detection to actually find stuff we're interested in. | So those connections are |
So those connections are | kind of a bit implicit here in the slides as we go from outlier detection to imbalance data. |
kind of a bit implicit here in the slides as we go from outlier detection to imbalance data. | Okay. |
Okay. | there's always a lot of benefits. So if you can do some data augmentation, you know, we can basically add training data, it's gonna reduce this imbalance and reduce some overfitting |
there's always a lot of benefits. So if you can do some data augmentation, you know, we can basically add training data, it's gonna reduce this imbalance and reduce some overfitting | and then these databases end up being pretty complicated to build. And |
and then these databases end up being pretty complicated to build. And | companies do not often share their comics. |
companies do not often share their comics. | Interesting, I think by that we actually mean companies don't often share their extra data. So sometimes the data out there, even in this. But you know your friend at the other company is not going to give it to you |
Interesting, I think by that we actually mean companies don't often share their extra data. So sometimes the data out there, even in this. But you know your friend at the other company is not going to give it to you | happens more often than I like to admit. |
happens more often than I like to admit. | Okay, so now we'll get to. |
Okay, so now we'll get to. | If we're gonna do this over sampling process. There's a couple of other ways. We can do data augmentation. There's a couple of nice ones we can definitely use for images that are pretty common |
If we're gonna do this over sampling process. There's a couple of other ways. We can do data augmentation. There's a couple of nice ones we can definitely use for images that are pretty common | where this is where we can basically take images and do some sort of like a fine transformation. You can make the image bigger, smaller. |
where this is where we can basically take images and do some sort of like a fine transformation. You can make the image bigger, smaller. | You can change the colors. You can change the contrast. You can just add Gaussian noise. But that's not typically how the world works. Usually it's just sort of like a change in perspective. Often it's darker or just like lighter outside. |
You can change the colors. You can change the contrast. You can just add Gaussian noise. But that's not typically how the world works. Usually it's just sort of like a change in perspective. Often it's darker or just like lighter outside. | So you can add those types of changes. So these are just some examples. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.