Quantization

Quantization is the process of constraining an input from a continuous or otherwise large set of values (such as the real numbers) to a discrete set (such as the integers).

Quantization for neural networks

At Ericsson, Adam Cooke told me about the project that he had been working on for over a year now on Bit Quantization.

The idea is to reduce from like 24 bit to 8-bit. Ericsson is working on a “DIY Tensorflow” library that works on EMCA. Adam shared with me some of the logic he was going through.

Dequantization