Развој кодера таласног облика за потребе неуронских мрежа и обраду сигнала

Abstract

This doctoral thesis aims to design low-bit scalar quantizers and analyze their application in Neural Networks (NNs) and signal processing. In this thesis, we consider the possibilities and limitations that rest on quantization, as a leading technique for data coding and compression. In particular, we examine the inevitable accuracy loss of signal and data presentation due to quantization in the signal processing area, as well as in many modern solutions, that use quantization. As stated in this thesis, there are a number of qualitative performance indicators, which indicate that appropriate quantizer parameterization can optimize the amount of data transmitted in bits. Quantized Neural Networks (QNNs) is a promising research area, especially important for resource constrained devices. Relying on a plethora of conclusions about scalar quantizers derived for signal processing tasks and taking into account the advantages of scalar quantization, we anticipate that by studying the statistical characteristics of neural network parameters, this thesis will contribute to determining an efficient weights compression solution utilizing new, well-designed scalar quantizers for post-training quantization.