tubetranscript.io

NPTEL IIT Guwahati · Transcript · · Updated

Lec 22: Generative AI – Transcript

Video by NPTEL IIT Guwahati · Watch the original on YouTube

Summary

AI-generated from the transcript. Verify details against the source video.

  • Generative modeling aims to learn the underlying probability distribution of a dataset in order to generate new, statistically consistent samples.
  • Maximum likelihood estimation is a fundamental statistical method used to determine the parameters that make the training data as probable as possible.
  • Generative models are categorized into explicit density models that evaluate probability density and implicit density models that learn mechanisms for sample generation.
Video thumbnail: Lec 22: Generative AI
Source video on YouTube · Transcript generated automatically and may contain errors.

Use this video with Pro

Reading, copying and transcript downloads are free. Create saved AI results for this video with a signed-in Pro account ($9/month).

Mindmap Pro · Sign in requiredTurn this transcript into a topic map you can review and download as PNG.Continue with this video →Flashcards Pro · Sign in requiredCreate questions and answers from this video to review and export.Continue with this video →X thread Pro · Sign in requiredCreate a source-based X thread draft to review before publishing.Continue with this video →Action items Pro · Sign in requiredFind explicit tasks and commitments mentioned in this video. Some videos have none.Continue with this video →

Transcript

Search ignores accents. All transcript text stays visible.

Hello and welcome to the lecture generative modeling. In this lecture, we first introduce generative learning models. Then we try to understand what are the statistical understanding for generative learning models. Then two important concept maximum acute and density estimation we discussed and then we study about different generative models.

Today we'll start generative modeling. Now as of now whatever we have discussed they are mainly discriminative learning or the models which follow the discriminative modeling.

What is the main way of doing discriminative modeling is that if we consider a vision task like a image classification,

we try to learn by discriminating ing some images. What is the example? If there is a cat image and we want to know that given a image is cat or non-cat

then we try to understand what factor or what attributes discriminate or discriminate between cat category or cat class and non-cat class. So when we are given a set of leveled or annotated examples,

we try to find out this mapping function. That function is essentially a discriminating function. And that's discriminating function is going to say that given a image is a cat image or a non-cat image and likewise.

So the discriminative learning basically follows a procedure. We are trying to find a hypothesis function or mapping function from a class of or a set of annotated example to a set of labels. So that by virtue of discriminating

attributes when a new instance is coming that we can say this is which class it belongs to.

So in general we do not try to understand what are the intrinsic or what is the statistical distribution of the given samples. Now we are going to a different paradigm of learning called generative modeling.

Now in this approach it mainly aims to learn the underlying probability distribution of a given data set.

Now that is the first fundamental difference between a discriminative learning and a generative learning or modeling that here we not only happy to find out the mapping function rather we aim to learn the underlying

probability distribution of our sample. or annotated sum or a given data set whatever is and what is the purpose after understanding the distribution our goal is subsequently generate new samples these new samples what we are generating

that are statistically consistent with the observed data observed data means given data That is the major agenda of generative modeling.

Now you can see here in this image this is a generator. So this is a neural net. Okay. from a random noise. So random noise is a latent space kind of.

Now this random noise let's say at Z P of Z means the jet space that using this generated the generator we generated a data distribution is called D bar for sale.

Now g of z is basically random noise and g is a generator. So we can get this data distribution. Now what is the intention is that given a data set this distribution what we want that these two distribution should be very

good. So we measure a statistical loss or lo statistical distance between D and D and we want to minimize this loss. So we have a a kind of this loss is back propagated to the this generative model so that it change its parameter such a way it more

similar sample can be generated. So now this is the setup. Now how it is consistent with this statement.

Now if you think it bit in different way that this is a observed data or a given data set real data distribution.

What actually we try to understand actually without knowing the data distribution we try to model this distribution in way how we are going to try to understand we are generating some

samples and when generating these two statistics these two data samples should be statistically closed.

So what what does this mean? We in a other way around we are trying to find this statistical or unknown statistical distribution by virtue of generating the similar samples. Once it has been trained this generative model or this generator without knowing the actual

we don't know this is a gshian mixture of gshian or what distribution the unknown distribution known by given the samples and but we model it that with the learn the underlying probability distribution.

This power distribution is being learned of unknown distribution so that it can generate the similar samples. So this is the basic idea of ch.

Okay. So let the training samples be drawn from unknown data distribution t. Now this is unknown data distribution. A generative model learns a parameterized approximation.

Y parameterized approximation. Now this is a parametric model this generator that is basically of Z of G of theta G and we what we want to do we

update this theta so that we can find the generator which generate more accur cubate samples or more similar samples real data distribution.

So parameterize approximation to this distribution by mapping samples from a simple Latin distribution like this typically randomized to a data space network.

So this is the basic idea of genetic model. Now we try to understand how this kind of model is being trained.

Now what is the our basic objective? As we have mentioned that basic objective is that this generator is a neural network or better to say a parameterized model neural model and that take the input from the random noise latins

Frequently asked questions

Discriminative learning tries to find a mapping function from annotated examples to labels, whereas generative modeling aims to learn the underlying probability distribution of the data to generate new statistically consistent samples.

Maximum likelihood estimation is a fundamental statistical method used to estimate the parameters of a generative model by maximizing the probability of observing the available training data.

Implicit density models do not explicitly calculate probability density, but instead learn a mechanism for generating samples, whereas explicit density models directly define and learn a probability distribution for the observed data.