Quantization for neural networks

The invention relates to a method and an apparatus for modifying a quantizer. In particular, at least one quantization level of a set of preliminary quantization levels is modified according to optimization of distortion involving a set of predetermined input values. At least one other quantization level in the set of preliminary levels is not modified. The quantization level that is not modified (non-modifiable) is either a minimum intercepted value or a maximum intercepted value. The modifications may help increase the dynamic range of quantized/dequantized data. This modified quantizer is advantageous for compressing data (e.g., feature maps, etc.) in neural networks. The quantizer can improve the accuracy of the neural network. 本发明涉及用于修改量化器的方法和装置。具体地，根据涉及一组预定输入值的失真的优化修改一组初步量化电平中的至少一个量化电平。不修改这一组初步电平中的至少一个其它量化电平。不被修改(不可修改)的量化电平是最小截取值或最大截取值。上述修改可以有助于增大量化/反量化数据的动态范围。这种修改后的量化器对于在神经网络中用于压缩数据(例如特征图等)是有利的。这种量化器可以提高神经网络的准确性。