INT8 quantization in PyTorch from scratch — scale factors, zero points, and integer arithmetic without QuantStub, to understand exactly what quantized layers do under the hood.

In this post, we will explore the Low Rank Approximation (LoRA) technique for shrinking neural networks for embedded systems. We will focus on the Convolutional Neural Network (CNN) case and discuss the rank selection process.

An elementary explanation of the problem with full rank matrices in neural networks and their solution via low rank approximations. Detailed explanation on how to set up the optimization problem and how to solve it, in possibly linear time.