Neural Network Quantization and Compression with Tijmen Blankevoort - TWIML Talk #292

In this discussion, Tijmen Blankevoort, a staff engineer at Qualcomm, delves into the fascinating world of neural network compression and quantization. He explains how much ML models can be compressed without losing efficiency and outlines the best strategies for achieving this. The conversation also touches on the lottery ticket hypothesis, exploring how feature selection can optimize neural networks. Tijmen reveals challenges in automating compression, like error propagation, and introduces innovative data-free quantization techniques that enhance performance across various models.