4. Data formats

As already noted, while learning phase calculations generally use 32-bit floats, inference calculations can make do with smaller formats.

It is difficult to establish rules for the data formats used in neural networks. This is because it is necessary to take into account the characteristics of the application, the desired precision of the results and the time required for inference. A number of publications have been published on data formats for learning. The article uses only +1 and –1 data. The