[Paper Review] Pruning Algorithms to Accelerate Convolutional Neural Networks for Edge Applications: A Survey

Jiayi Liu, Samarth Tripathi+2 more|arXiv (Cornell University)|May 8, 2020

Advanced Neural Network Applications52 references41 citations

TL;DR

A comprehensive survey of CNN pruning techniques for edge deployment, covering criteria, procedures, and trade-offs, with a discussion of practical challenges and future directions.

ABSTRACT

With the general trend of increasing Convolutional Neural Network (CNN) model sizes, model compression and acceleration techniques have become critical for the deployment of these models on edge devices. In this paper, we provide a comprehensive survey on Pruning, a major compression strategy that removes non-critical or redundant neurons from a CNN model. The survey covers the overarching motivation for pruning, different strategies and criteria, their advantages and drawbacks, along with a compilation of major pruning techniques. We conclude the survey with a discussion on alternatives to pruning and current challenges for the model compression community.

Motivation & Objective

• Motivate the need for model compression to enable edge/IoT deployment of CNNs.
• Characterize pruning criteria (data-agnostic vs data-driven) and their trade-offs.
• Describe pruning workflows and iterative training strategies for effective pruning.
• Summarize hardware-aware considerations and practical limitations of pruning in edge settings.

Proposed method

• Classify pruning criteria into data-agnostic and data-driven (data-aware) categories.
• Describe iterative pruning workflows combining pruning, masking, and retraining (Algorithm 1 lineage).
• Review saliency measures and optimization approaches (L1/2 norms, Hessian-based, APG, ADMM, etc.).
• Discuss structured vs unstructured pruning and the move toward structure-aware pruning for speedups.
• Summarize learning-based and RL/NAS approaches to automate pruning and compression budgets.

Experimental results

Research questions

• RQ1 What pruning criteria best identify non-critical neurons or filters without sacrificing accuracy?
• RQ2 How do iterative pruning procedures balance pruning intensity with retraining to recover accuracy?
• RQ3 What are the practical hardware-related gains (FLOPs, parameters, real speed) when pruning CNNs for edge devices?
• RQ4 What are the primary trade-offs and limitations of pruning versus alternative compression methods for edge deployment?
• RQ5 What future directions and open challenges remain in pruning for edge AI?

Key findings

• Pruning can reduce model size and inference time by substantial margins, but speedups do not always correlate directly with FLOPs reductions on real hardware.
• Structured pruning (e.g., channel/filter/pruned blocks) tends to yield more realistic speedups than unstructured prune-by-weight, due to better hardware utilization.
• Data-aware criteria (e.g., APoZ, activation entropy, neural importance scores) have grown in prominence for better preserving accuracy during pruning.
• Iterative pruning with retraining generally outperforms one-shot pruning, though the optimal training-from-scratch vs. retrain-after-prune results can vary by model and budget.
• There is a need for hardware-aware and energy-aware pruning paradigms to capture true resource utilization on edge devices.