Hierarchical Reinforcement Learning for Neural Network Compression (HiReLC): Pruning and Quantization
Kamar Hibatallah Baghdadi, Kawther Guoual Belhamidi, Sara Belhadj, Aissa Boulmerka, Nadir Farhi
Read on arXiv →
Key claim
Achieves up to 6.72x compression with minimal accuracy loss.
In plain English
Imagine you're trying to make deep neural networks smaller and faster without losing too much accuracy. This is a common challenge because as you reduce the size of a model, you often end up sacrificing performance. Current methods can struggle with this balance, leading to either excessive accuracy loss or inefficient compression strategies. This is what's called the trade-off problem in model compression.
What this paper introduces is a new way to tackle that trade-off. Instead of just compressing a model in one go, it breaks the process into two levels: low-level agents focus on specific parts of the model, like individual layers or blocks, while high-level agents manage the overall strategy. This means that each part of the model can be optimized independently, which helps in making smarter decisions about how to compress without losing too much accuracy.
The key result is that this approach can achieve compression ratios of nearly 6 times while only experiencing minor accuracy drops in most cases. This is a significant improvement over traditional methods, making it a valuable tool for anyone looking to deploy efficient deep learning models in real-world applications.
The framework introduces a novel hierarchical approach to neural network compression that combines reinforcement learning with sensitivity analysis.
The experiments demonstrate effective compression ratios and accuracy impacts across multiple benchmarks, supporting the claims made.
Deep reliability assessment
The methodology supports the claim that hierarchical, sensitivity-guided RL can find roughly 6× effective parameter-storage compression policies across several ViT and CNN image-classification settings with modest accuracy loss after fine-tuning. It does not strongly support claims about deployment efficiency, because latency, energy, backend packing overhead, and matched compression baselines are not fully evaluated.
Reproducibility
No open-source code or repository is mentioned in the provided paper text. The experiments use public datasets such as CIFAR-10, CIFAR-100, and Tiny ImageNet, and one DeiT-Small/CIFAR-100 setting reports multi-seed results over seeds 42, 123, and 456, but full reproduction would require implementation details for the RL environment, compression operators, surrogate loop, and fine-tuning setup.
Key figure
The key architecture is a two-level controller where high-level ensemble agents allocate global compression budgets using Fisher sensitivity, while per-block low-level agents choose bitwidth, pruning keep-ratio, quantization type, and granularity, with a surrogate-guided active learning loop reducing evaluation cost.