Evolutionary perspective on model fine-tuning

Andrei Kucharavy, Ljiljana Dolamic, Rachid Guerraoui

Keywords: Evolutionary algorithms, stochastic gradient descent, fine-tuning

Abstract:

Be it in natural language generation or in the image generation, massive performances gains have been achieved in the last years. While a substantial part of these advances can be attributed to improvement in machine learning architectures, an important role has also been played by the ever-increasing parameter number of machine learning models, which made from-scratch retraining of the models prohibitively expensive for a large number of users.
In response to that, model fine-tuning - starting with an already good model and further training it on the data relevant to a new, related problem, gained in popularity. This fine-tuning is formally similar to the natural evolution of genetic codes in response to shifting environment.
Here, we formalize this similarity in the framework of Fisher Geometric model and extreme value theory and present a set of tricks used by naturally evolving organisms to accelerate their adaptation, applicable to model fine-tuning.

One-sentence Summary: We formalize a parallel between biological systems evolution and machine learning models fine-tuning to suggests ways of accelerating fine-tuning