Quantization of weights in capsule neural networks

The influence of the weight quantization levels number on a capsule neural network functioning quality is studied. The network is tested on the MNIST dataset recognition. It is shown that to achieve the performance of the network with continuous weights it is enough to have 16 levels of uniform quantization and 8 levels of exponential quantization. The quantized weights of the capsule neural network can be successfully implemented based on multilevel memristors.