A comprehensive survey on model quantization for deep neural networks in image classification

Analysis

Why This Paper Matters

As deep neural networks achieve state-of-the-art accuracy in image classification, their growing size and computational demands pose significant barriers to deployment on edge devices, mobile phones, and embedded systems. Model quantization has emerged as a critical technique to reduce memory footprint and inference latency while preserving accuracy. This comprehensive survey fills a gap by systematically organizing the vast landscape of quantization methods, making it an essential resource for AI practitioners seeking to optimize models for real-world applications.

The paper's timing is particularly relevant given the industry's push toward on-device AI and the proliferation of large-scale models. By providing a structured overview of uniform, non-uniform, and mixed-precision quantization, it helps practitioners navigate the trade-offs between compression ratio, accuracy retention, and hardware support. The survey also highlights practical considerations such as calibration data requirements and training overhead, which are often overlooked in method-focused papers.

Technical Contributions

• Taxonomy of Quantization Methods: The paper categorizes techniques into post-training quantization (PTQ) and quantization-aware training (QAT), further subdividing by bit-width (e.g., 8-bit, 4-bit, binary) and granularity (per-layer, per-channel, per-group).
• Analysis of Accuracy-Efficiency Trade-offs: It systematically compares methods on standard benchmarks (e.g., ImageNet, CIFAR-10), showing that 8-bit quantization typically yields <0.5% accuracy drop while 4-bit methods may incur 1-2% degradation without advanced techniques like clipping or fine-tuning.
• Hardware Considerations: The survey discusses compatibility with common hardware (GPUs, CPUs, FPGAs, ASICs) and notes that uniform quantization is more hardware-friendly than non-uniform or mixed-precision schemes.
• Future Directions: Identifies open challenges including handling of activation outliers, quantization for transformers, and automated mixed-precision search.

Results

While the survey does not present new experimental results, it aggregates findings from prior work: 8-bit quantization achieves near-lossless accuracy (within 0.1-0.5% of full-precision) on ImageNet for ResNet and MobileNet architectures. 4-bit quantization with QAT can maintain accuracy within 1% of baseline, while binary quantization often suffers 5-10% drops. Compression ratios range from 4x (8-bit) to 32x (binary), with corresponding speedups of 2-4x on specialized hardware.

Significance

This survey provides a foundational reference for the growing field of efficient deep learning. By clarifying the landscape of quantization techniques, it enables practitioners to make informed decisions when deploying models in resource-constrained environments. The paper's emphasis on hardware-aware methods and mixed-precision quantization aligns with current trends in automated machine learning and edge AI. As models continue to grow in size, quantization will remain a key enabler for democratizing AI across devices, and this survey helps accelerate progress by consolidating knowledge and highlighting promising research directions.