Method for establishing data-recognition model

A computer system for establishing a data-recognition model, comprising:

• a data storage (1) that stores a deep neural network (DNN) and a set of a number (X) of dithering algorithms, where X≥2;
• and
• a processor (2) coupled to said data storage (1) and programmed to implement the steps of:
• A) generating a number (Z) of Y-combinations of dithering algorithms from the set of the number (X) of dithering algorithms, each of the Y-combinations including a number (Y) of dithering algorithms, where 1≤Y≤(X−1) and the number (Z) is equal to xCy;
• B) for each of the number (Z) of Y-combinations of dithering algorithms, using the number (Y) of dithering algorithms of the Y-combination to perform a dithering operation on a to-be-processed data group represented in (a) number of bit(s), so as to obtain, in total, a number (Z) of size-reduced data groups each being represented in (b) number of bit(s), where 1≤b≤(a−1);
• C) performing a number (Z) of training operations on the DNN using the number (Z) of size-reduced data groups, respectively, so as to generate, for each of the number (Z) of training operations, a DNN model, a training result of the training operation, and a steady deviation between the training result and a predetermined expectation;
• and
• D) selecting one of the number (Z) of Y-combinations of dithering algorithms corresponding to the size-reduced data group that results in the training result with the smallest steady deviation as a filter module, selecting the corresponding DNN model as the data-recognition model, and generating layout of logic circuits hardware corresponding to the established data-recognition model.