Hattie Zhou: Lottery Tickets and Algorithmic Reasoning in LLMs

Fortuitous Forgetting

The concept of fortuitous forgetting in neural networks explores how forgetting can aid learning and generalization. Hattie Zhou explains that, similar to humans, forgetting undesirable information can help models relearn and improve performance 1. This iterative process involves removing unwanted data and retraining the model to enhance desirable traits. Zhou notes that this approach can lead to faster grokking, a phenomenon where models suddenly improve their test accuracy after prolonged training 2.

The hypothesis is that when you do this, every time you retrain, you're starting from a better initialization that has less of the undesirable information and more of the desirable information.

--- Hattie Zhou

This method aligns with other neural network phenomena like double descent, where models simplify over time to improve generalization.

   

Iterative Processes

Iterative training processes are crucial for enhancing model performance and understanding learning dynamics. Hattie Zhou discusses how these methods, though expensive, can be effective in refining models by repeatedly improving upon previous iterations 3. She highlights examples where models generate labels, fine-tune on them, and iteratively enhance their performance. This approach is seen as a paradigm for model improvement, especially in large language models.

The iterative process is kind of taking something that works and just multiplying it and almost bootstrapping the process, and it's easy.

--- Hattie Zhou

Additionally, Zhou mentions that during training, weights may converge to zero but later become significant, indicating the complex dynamics of learning and forgetting 4.