Image Classification Learning Based On Quantized Multilayer Perceptron

Neural networks are widely used in computer vision,natural language processing,speech recognition and other fields,which have achieved excellent results.However,because neural networks occupy a lot of storage space and energy in the training and inference process,it is difficult to deploy on intelligent mobile devices or embedded systems with limited resources.In order to reduce model size and energy consumption,parameter quantization is considered to be a very effective method.Therefore,with the development of intelligent mobile devices,neural networks with discrete parameters have become a research hotspot in recent years,and more and more scholars begin to study new quantitative schemes and network training methods.Taking image classification task as an example,this paper studies a quantized neural network with one hidden layer.The networks contain two weight connections.According to the different feature extraction methods,two quantization network models based on block image transformation and convolution feature are discussed in this paper.For quantization classification network based on block feature transformation,this paper uses traditional block image transformation and a special structure matrix operation to extract pixel information and spatial position information respectively.We then introduce a random label to simplify the network into a nonlinear quantized plus linearly connected single-layer network.In the training phase,the traditional backward propagation algorithm is used to learn the designed special matrix.In the inference phase,a new classification method based on k nearest neighbor method is proposed in this paper,and its performance is proved to be better than kNN through experiments.For the quantitative classification network based on convolutional features,the update of weights in the first layer is modeled as a discrete compressed sensing.Firstly,target propagation is carried out for this layer,and then the weights of this layer is learned by using the reconstruction algorithms of discrete compressed sensing to make the output of this layer match the propagated target before.Then,based on sign consistency,two discrete compressed sensing reconstruction algorithms are proposed innovatively,namely,maximum coherence value algorithm and message passing algorithm.Finally,we verify the performance of the reconstruction algorithm and its application to the classification of neural networks through experiments.