August 2022 Summaries

Allen AI's recent paper introduces a technique called layer recycling, which caches outputs from certain intermediate layers during training and inference phases, resulting in an 83% speedup with minimal performance loss, specifically for language models. This method accelerates processes by avoiding repeated data passes through frozen layers, caching outputs instead for future epochs, although it is task-dependent and more effective in larger models or lower-end machines. Experiments conducted with Quaterion, which extends this caching concept to various data types via an intelligent key extractor, reveal that while 50% layer recycling yields performance close to full training, task-specific adaptations and dataset sizes significantly impact results. Findings indicate that smaller datasets exacerbate performance drops in full training and layer recycling, yet training only the EncoderHead is more resilient under these conditions. The variation in performance across tasks underscores the need for further research, and the flexibility of Quaterion facilitates such experimentation.