Get the latest tech news
Diffusion models are evolutionary algorithms
Authors: Yanbo Zhang, Benedikt Hartl, Hananel Hazan, Michael Levin
Imagine we have a complex task (for example, finding the optimal shape of an airplane wing), and we create a set of random solutions – like a "population" of creatures in nature. Many biologically inspired evolutionary algorithms can be understood similarly: they optimize an objective function by maintaining and iteratively changing the distribution of a large population. In evolutionary algorithms, fitness evaluation is often the most computationally expensive operation, so the authors tried to reduce the number of iterations by borrowing cosine scheduling from work on diffusion models.
Or read this on Hacker News
