Layerwise Quantization: A Replication Study

2025 #research

Replication of Chen et al.'s layer-wise DNN quantization paper on CIFAR-10 and TinyImageNet, confirming results hold under limited data conditions.

View project →

This was a replication study of “Deep Neural Network Quantization via Layer-Wise Optimization Using Limited Training Data” by Shangyu Chen et al. The motivation is practical: deep learning models are powerful but expensive to run, and quantization is one of the more principled ways to compress them for low-compute devices.

Rather than replicating on ImageNet (the original paper’s benchmark), I used CIFAR-10 and TinyImageNet — smaller but sufficient to test whether the core claims hold under more limited data conditions. They do. Layer-wise quantization achieves performance comparable to full-precision models with minimal degradation, even when the training data is constrained.

The full report is linked above.