Types of Generative AI models
The most significant advances in the field of artificial intelligence has definitely been Generative AI.
This enables users to generate new content based on a variety of inputs.
Generative AI models use neural networks to identify the patterns and structures within existing data to generate new and original content
The prominent frameworks involved with Generative Artificial Intelligence are mentioned below :
i. Diffusion models
ii. Variational auto-encoders (VAEs)
iii. Generative adversarial networks (GANs)
iv. Transformer networks
I. DIFFUSION MODELS
Imagine we have a image and we add some gaussian noise to it, making it a bit blur. Then we do this again and repeat it several times.
Eventually, we will have an unrecognisable picture of static sample of pure noise.
Now, what if we could figure out how to undo the process ?
i.e. begin with the noise image and get to the original image, by gradually removing the noise and end up with a coherent image.
This is the basic idea behind Diffusion models.
An approach gaining traction in generative modelling, particularly used in Image generation.
Diffusion models have also been known to outperform generative adversarial networks (GANs) in perception quality metrics. They have also known to show impressive performance in conditional settings such as converting text descriptions to images or painting and manipulation.
II. VARIATIONAL AUTO-ENCODERS (VAEs)
VAEs consist of two neural networks : encoder and decoder.
The encoders in VAEs optimise for more efficient ways of representing data, whereas the decoders optimise for more efficient ways of regenerating the original data set. The encoder and decoder work together to learn an efficient and simple latent data representation.
VAEs are the most effective and useful process for creating generative models.
III. GENERATIVE ADVERSARIAL NETWORKS (GANs)
Let’s break the term GAN into 3 parts “Generative-Adversarial-Network”.
Generative : Generates data (creates fake data)
For e.g. Given a task to generate the image of MonaLisa, the GANs creates an outline for the image of MonaLisa.
If the image was built on a model which was trained on more epochs.
The point of “Generative” is it generates data that is an exact representation of the real data.
Adversarial : This part has two components “Generator” and “Discriminator”.
The “Discriminator” is like our machine learning algorithm. Lets say we want to classify an animal as cat or dog, so that is a discriminator which means we are trying to discriminate the animals between the two categories.
Think of Generator and Discriminator competing to win, Generator tries to create fake data and Discriminator prevents in creating fake data; basically discriminator helps the generator in creating better and more accurate results by discriminating the data points which are not relevant.
Networks : Networks can be Deep convolutional or fully connected.
IV. TRANSFORMER NETWORKS
Transformers are models that can translate text and even generate computer code. BERT, GPT-3, T5 all of these models are based on Transformers.
They are designed to process sequential input data non-sequentially.
Transformers prove to be extremely powerful in performing advanced Natural language processing tasks.
This was a short articles on knowing about the types of Generative AI.
In the upcoming articles, we will understand these topics in detail.
Stay Tuned!
