Optimizing Deep Learning Models for Resource-Constrained Environments With Cluster-Quantized Knowledge Distillation

Deep convolutional neural networks (CNNs) are highly effective in computer vision tasks but remain challenging to deploy in resource-constrained environments due to their high computational and memory requirements. Conventional model compression techniques, such as pruning and post-training quantization, often compromise model accuracy by decoupling compression from training. Furthermore, traditional knowledge distillation approaches rely on full-precision teacher models, limiting their effectiveness in compressed settings. To address these issues, we propose Cluster-Quantized Knowledge Distillation (CQKD), a novel framework that integrates structured pruning with knowledge distillation, incorporating cluster-based weight quantization directly into the training loop. Unlike existing methods, CQKD applies quantization to both the teacher and student models, ensuring a more effective transfer of compressed knowledge. By leveraging layer-wise K-means clustering, our approach achieves extreme model compression while maintaining high accuracy. Experimental results on CIFAR-10 and CIFAR-100 demonstrate the effectiveness of CQKD, achieving compression ratios of 34,000× while preserving competitive accuracy—97.9% on CIFAR-10 and 91.2% on CIFAR-100. These results highlight the feasibility of CQKD for efficient deep learning model deployment in low-resource environments.