Model training

Generative AI Concepts

Daniel Tedesco

Data Lead, Google

Pick your mode of train-sportation

Image of a plane and a car.

  1. Hardware
  2. Time
    • Dataset size
    • Model complexity
    • Rounds of training
  3. Cost
Generative AI Concepts

Graduate to advanced techniques

  • Foundation generative AI model is a first step
  • Advanced techniques specialize them for specific contexts

Key advanced techniques:

  1. Transfer learning and fine-tuning
  2. RLHF
  3. Custom embeddings

A picture of students at their graduation

Generative AI Concepts

From cats to lions

  • Transfer learning transfers knowledge from one task to another
  • Fine-tuning is a type of transfer learning for a small dataset

A model learning from lion images, saying "Still furry, four paws... I got it!", and then producing a lion image

1 Creative Commons Attribution-Share Alike 4.0, Bing image generator
Generative AI Concepts

Where does your feedback go?

Image of the Google Bard chat interface with the 'thumbs up' and 'thumbs down' feedback options highlighted

1 Google Bard, 2023
Generative AI Concepts

Thumbs up for better responses

Several text replies rated thumbs up or thumbs down being passed to a robot that says, "Okay, now I know what kind of answers the humans like"

1 Google Bard, 2023
Generative AI Concepts

Embeddings as fingerprints

  • Similar to recognizing a fingerprint
  • Unique representations of data entities
  • Capture meaning, context, and relationships in compact form

A series of fingerprints

1 Wikimedia commons
Generative AI Concepts

Embedding Dan

Various portrait pictures of Daniel Tedesco being given to a generative AI

1 Daniel Tedesco, 2023
Generative AI Concepts

Embedded Dan

A model producing a variety of stylized output headshots of Daniel Tedesco

1 Daniel Tedesco portraits from vana.com
Generative AI Concepts

Let's practice!

Generative AI Concepts

Preparing Video For Download...