Bismillah

Over the past few years, there's been a lot of buzz on Generative AI products and projects such as Open AI's Sora and DALL.E.

A question that some of us have but don't bring up often is, "What is Generative AI?" We can look at today's examples of Generative AI projects and have a rough idea of what it is, but can we go deeper to understand it more? Well, let's try!!

What's Generative AI?

Simply put, this AI can create(generate) new things(yes, you're right!). They do this through different means as will be seen in the different types of Generative AI models.

In machine learning, the models used for this type of AI fall under the Unsupervised Learning category/classification.

To bring this closer to home, models can also be classified in two ways:

1. Discriminative Models: this is seen in how most Machine Learning Models classify data/input or predict values in their output.

2. Generative Models: this type of Machine Learning Model works by attempting to judge if the data it receives has been seen before. Using this "knowledge" and the patterns it observed, the model can then be used to create(generate) new examples of data close to what was presented to it.

With that in mind let's look at some types of Generative AI models.

Types of Generative AI

Task-Specific Generative AI Models

1. Auto-Regressive Models

   These models use Autoregressive Convolutional Neural Networks(AR-CNNs) for studying systems that evolve with time and assume that the likelihood of some data is solely dependent on past events.

2. Generative Adversarial Networks.

   Just like the name suggests, the goal of these models is to pit two "adversary" networks against each other to generate new content.

   The two networks pitted against each other by the training algorithm are:

   • Generator Network(Generator): This produces/generates new data.

   • Discriminator Network: This measures/judges how closely the generator's data represents the training dataset.

The goal of this back and forth between the networks is to make both networks better through the feedback loop and thus more accurate output is expected.

3. Transformer-based Models

   These models are common in natural language modeling as they are used to study sequentially structured data.

4. Diffusion Models

   These models can create new data by utilizing the training data. For instance, Generating faces from examples of human faces in training data.

   They work by transforming simple and easily obtainable distributions into more complex and meaningful data distributions.

5. Variational Autoencoders(VAEs)

   These models obtain a compact representation of data by combining the powers of probabilistic modeling and autoencoders. They sample points from the obtained distribution, to create fresh samples by encoding input data into a lower-dimensional latent space.

6. Flow Models

   These models process the probability distribution of different values or events in a dataset to learn the underlying structure. With this distribution, it can then generate fresh data points, maintaining the initial dataset's characteristics and statistical properties.

Checkpoint: Type 2 - General AI Models

At this point I do understand that it gets more complex, so let's discuss the different types of Generative AI's in the coming blog post.

See you soon.

Asalaam Aleykum!!