Evolution Strategies at the Hyperscale
13 days ago
- #Machine Learning
- #Evolution Strategies
- #Neural Networks
- Introduction of EGGROLL, an evolution strategies (ES) algorithm for large neural networks.
- EGGROLL uses low-rank matrix perturbations to reduce computational and memory costs.
- Theoretical analysis shows EGGROLL converges to full-rank updates at a fast rate.
- Experiments demonstrate EGGROLL's performance in RL, LLM reasoning, and pre-training of recurrent language models.