HASTE: A Framework for Training-Free, Dynamic, and Steerable Compression of Pre-Trained Convolutional Neural Networks

2026-06-29

AI Summary

This work addresses the challenge of deploying large convolutional neural networks (CNNs) on resource-constrained devices, where high computational costs are prohibitive and existing dynamic compression methods often require retraining or data access, limiting their applicability to pretrained models. To overcome this, the authors propose HASTE—a plug-and-play convolutional module that enables training-free, dynamic, and controllable compression of pretrained CNNs for the first time. HASTE leverages locality-sensitive hashing (LSH) during inference to dynamically identify and merge redundant channels in feature maps, simultaneously compressing both input features and the corresponding convolutional kernels along the depth dimension to substantially reduce FLOPs. Experiments demonstrate that HASTE reduces FLOPs by 46.2% on ResNet-34 for CIFAR-10 with only a 1.25% accuracy drop, and its effectiveness and generalizability are further validated across ImageNet and diverse network architectures.

Abstract

Deploying large convolutional neural networks (CNNs) on resource-constrained devices is challenging due to their high computational cost. While dynamic execution methods are promising, existing approaches for CNNs typically require specialized training or fine-tuning, limiting their effectiveness when applied to pre-trained models and requiring data access. To address this gap, we propose HASTE (Hashing for Tractable Efficiency), a plug-and-play convolution module that enables training-free, dynamic compression of large pre-trained CNNs. At inference time, HASTE uses locality-sensitive hashing to identify and merge redundant channels of latent feature maps on a patch-wise basis. This process simultaneously compresses the depth of both input features and their corresponding filters, resulting in computationally cheaper convolutions. We conduct extensive experiments on CIFAR-10 and ImageNet across a range of architectures, demonstrating a 46.2% FLOPs reduction in a ResNet34 on CIFAR-10 with only a 1.25% drop in accuracy, without any retraining. We support our claims by comprehensive ablation studies to validate our core design choices, an analysis of the method's properties and limitations, and a discussion that connects our channel merging scheme to the conceptually related task of token merging in Vision Transformers. Our results demonstrate that HASTE provides an effective solution for steerable compression of pre-trained CNNs at runtime, opening new possibilities for the deployment of efficient deep learning methods.

Problem

• dynamic compression
• pre-trained CNNs
• training-free
• resource-constrained deployment
• computational efficiency

Innovation

• training-free compression
• dynamic channel merging
• locality-sensitive hashing
• plug-and-play module
• steerable inference