Hands-On Generative AI with Transformers and Diffusion Models

Chapter 3. Compressing and Representing Information

This chapter introduces ML models and techniques to learn efficient data representations for tasks involving images, videos, or text. Why are efficient representations important? We want to reduce the amount of information we need to store and process while keeping the essential characteristics of the data. Rich representations enable training models specialized on particular tasks, and making the representations compact reduces the computational requirements to train and work with data-intensive models. For example, training on a vector embedding of an image can be more efficient and expressive than doing it directly on its pixels.

Traditional compression methods like ZIP or JPEG focus on specific data types and use handcrafted algorithms to reduce file sizes. While these methods are effective for their intended purposes, they lack the flexibility and adaptability of learned compression techniques. ZIP, for instance, excels at lossless compression of general data by identifying and encoding repetitive patterns. On the other hand, JPEG is designed specifically for image compression and achieves significant size reduction by discarding less noticeable visual information. However, these traditional methods don’t learn from the data they compress and can’t automatically adapt to different types of content or optimize for specific tasks beyond size reduction. This is where ML models can be useful.

We’ll begin by exploring AutoEncoders, ...