MACHINE LEARNING - AI Dynamics

Why does GM work? The trick is to make everything linear: Train generators, which are linear operators, to fulfill the Kolmogorov equation, a linear equation. Then, marginalizing across data points is simple leading to training methods that work on large datasets.

→ View original post on X — @mit_csail

8 November 2024

Markov Processes Parameterization Unifies Multimodal GenAI Models

By

@mit_csail

–

08 November 2024 18h17

Many GenAI models are based on Markov processes. By parameterizing them w/"generators," they can:
1. Unify & universally characterize Markov models
2. Train them at scale
3. Combine them to build multimodal models or superpositions
4. Train new models, e.g. jump models

→ View original post on X — @mit_csail

8 November 2024

FrameJump: Novel Jump Models for Protein Generation

By

@mit_csail

–

08 November 2024 18h17

Their experiments show:
1. Jump models are an unexplored model class w/promising results.
2. Combining models via Markov superpositions improves generation results. Check out FrameJump, a novel protein generation method — they can make amino acids jump to a protein! pic.twitter.com/vUNm2XxHs6
— MIT CSAIL (@MIT_CSAIL) 8 novembre 2024

Their experiments show:
1. Jump models are an unexplored model class w/promising results.
2. Combining models via Markov superpositions improves generation results. Check out FrameJump, a novel protein generation method — they can make amino acids jump to a protein!

→ View original post on X — @mit_csail

8 November 2024

Unified Theory of Generative AI: Generator Matching Framework

By

@mit_csail

–

08 November 2024 18h17

Can we design a generative AI algorithm that works for any data type? Together with a team at FAIR, CSAIL researchers propose a "unified theory of GenAI" called "Generator Matching" (GM). GM brings together a range of GenAI models, allowing researchers to build models for any

→ View original post on X — @mit_csail

8 November 2024