Power-of-Two Quantization for Low Bitwidth and Hardware Compliant Neural Networks

Power-of-Two Quantization for Low Bitwidth and Hardware Compliant Neural Networks

Explore power-of-two quantization techniques for low bitwidth and hardware-compliant neural networks in this 24-minute conference talk from the tinyML Research Symposium 2022. Presented by Dominika Przewlocka-Rus, a researcher at Meta Reality Lab Research, the talk covers key problems in quantization, various quantization methods, and their key differences. Learn about straight-through estimation, examine results, and consider hardware implications. The presentation concludes with a Q&A session and acknowledgment of sponsors, providing valuable insights for those interested in optimizing neural networks for resource-constrained environments.