Why this Class?
- Deep learning is the holy grail for almost all modern AI tasks
- Breakthroughs with deep leanring
- LLMs (Large Language Models) - ChatGPT, Llama-2, PaLM-2
- Diffusion model - Stable Diffusion, Midjourney
- Multimodal: DALL-E 3
- Automatic speech recognition
Large Language Model (LLM)
- Probabilistic model that can generate probabilities of a series of words, based on text corpora it was trained on
- Built by a series of Transformers
- GPT-3 (175B)
- LLama-2 (7B, 13B, 70B)
- PaLM-2 (rumor: ~340B (L))
- GPT-4 (rumor: > 1,000B)
- It can perform
- Reasoning
- Multilingual translation
- Coding
Diffusion Model
- An image generative model
- Allows you to sample from an image distribution
- Unconditional / conditional
- Stable Diffusion, Midjourney are all text-to-image models
Flaws / Dangers of AI
Although Deep Learning Models have made huge strides in recent years, they still contain many flaws, we are still far from an AGI.
Hallucinations: LLMs will sometimes generate information from nowhere, providing misleading information. i.e generating a fake city name.
Vulnerability To Input Perturbations: AI’s can output incorrect responses if inputs are designed to perturb it. i.e adversarial pattern causing Panda to be classified as a Gibbon. More dangerously adversarial patterns on street signs or jailbreaking LLMs.
DALL-E 3
DALL-E 3 combines ChatGPT with image generative model for better understanding and faithful generations, link to video.