Quantization consists of replacing the original data types used to represent the weights of neural networks with less resource-intensive data types. Much work has been done on quantization, however, existing methods that provide guarantees simply give error bounds on the difference between the original and reduced precision.

In this article, we propose a new quantization technique that instead of guaranteeing error bounds finds the minimum precision required to maintain dominance (independent of a given set of formats). This means that, independently of what the scores of each class are, we guarantee that the dominant class remains the same in the original and quantified networks.

Our method is static and the proposed quantization holds for all the inputs. Technically, we use existing theorems that give error bounds on dot products, and we derive an optimization problem whose solution gives the reduced precision. Experimental results are presented.