Hierarchical Reinforcement Learning for Neural Network Compression (HiReLC): Pruning and Quantization

Researchers introduce HiReLC, a hierarchical reinforcement learning framework that automates the joint compression of neural networks through pruning and quantization. The system achieves 5.99-6.72x compression ratios across Vision Transformers and CNNs with minimal accuracy loss, using a two-level agent architecture guided by Fisher Information sensitivity estimates.

HiReLC addresses a critical challenge in neural network deployment: reducing model size and computational requirements without sacrificing performance. The framework's innovation lies in its hierarchical decomposition, where low-level agents optimize individual network blocks while high-level agents coordinate global resource allocation through ensemble voting. This architectural approach sidesteps the computational explosion inherent in searching compression configurations across entire networks simultaneously.

The broader context involves the ongoing tension between model capability and practical deployment. As deep learning models grow increasingly large—particularly Vision Transformers—their computational and memory demands become prohibitive for edge devices, mobile platforms, and resource-constrained environments. Previous compression methods typically treated pruning and quantization as sequential or independently optimized tasks, often yielding suboptimal results. HiReLC's joint optimization represents a maturation of compression research methodology.

The practical implications extend across multiple sectors. For computer vision applications, edge AI deployment, and mobile inference, achieving 6x compression with 0-3.83% accuracy variance enables deployment scenarios previously infeasible. The architecture-agnostic design strengthens its applicability—the modular abstraction means the controller generalizes across different network topologies without redesign.

The integration of active learning and surrogate models demonstrates sophisticated research engineering; using lightweight MLP surrogates to guide policy optimization rather than replace final evaluation balances computational efficiency with empirical rigor. Moving forward, the key metrics to monitor include reproducibility across additional architectures, scalability to state-of-the-art large language models, and real-world inference speedups on actual hardware.

• →Hierarchical RL framework achieves 5.99-6.72x neural network compression with minimal accuracy degradation across Vision Transformers and CNNs
• →Two-level agent design optimizes both local block-level configurations and global budget allocation through Fisher Information-guided sensitivity analysis
• →Architecture-agnostic controller with modular layer abstraction enables generalization across different network topologies without framework redesign
• →Active learning loop combining surrogate-guided optimization with post-compression fine-tuning reduces computational cost of policy evaluation
• →Joint quantization and pruning search over multi-discrete action spaces outperforms sequential or independent compression approaches