Advances in Neural Network Pruning and Architecture Search
Recent developments in the field of neural network optimization have seen significant advancements in pruning techniques and neural architecture search (NAS). The focus has been on improving computational efficiency, reducing model size, and enhancing generalization capabilities, particularly under distribution shifts.
Pruning Innovations: The field has seen a shift towards more structured and gradient-aware pruning methods, which aim to maintain or even improve model accuracy while significantly reducing the number of parameters. These methods often leverage reinforcement learning or fixed-rate strategies to determine optimal pruning distributions across network layers, ensuring that the most critical components are preserved.
NAS Developments: NAS has progressed towards more efficient and data-free evaluation methods, with a focus on reducing the computational overhead associated with architecture performance evaluation. Techniques such as zero-shot NAS and gradient-based search mechanisms are being employed to accelerate the search process and improve the quality of selected architectures.
Noteworthy Papers:
- FGGP: Fixed-Rate Gradient-First Gradual Pruning - Introduces a novel gradient-first pruning strategy that outperforms state-of-the-art methods in various settings.
- Zero-Shot NAS via the Suppression of Local Entropy Decrease - Proposes a data-free proxy for architecture evaluation, significantly reducing computation time while maintaining high accuracy.
- RL-Pruner: Structured Pruning Using Reinforcement Learning - Utilizes reinforcement learning to optimize pruning distribution, achieving effective model compression and acceleration.
- Ghost-Connect Net: A Generalization-Enhanced Guidance For Sparse Deep Networks Under Distribution Shifts - Enhances network adaptability to distribution shifts by introducing a companion network for connectivity-based pruning.