Approximated deep neural networks : A computationally efficient inference

Abstract

For most of the pattern recognition and machine learning tasks, deep learning is performing well in recent years. Solving these machine learning tasks require large size deep neural networks (DNNs). Many state-of-the-art DNNs consists of millions of training parameters and arithmetic operations. Increase in size of deep neural network models results in high computational and memory space needs. Therefore, realization of these models on low power devices is possible with some approximation while retaining the accuracy of network. Many recent works have proposed different data quantization techniques. Most of these works require retraining of DNNs after quantization for reducing the accuracy loss due to quantization. Other works which do not retrain the network after quantization, suffer loss in accuracy. This thesis presents a scalable technique for representing trained parameters of deep neural networks in such a way that these trained parameters can be used for computational and memory-efficient inference phase of deep neural networks. This technique consists of two variations i.e. log 2 lead, and ALigN . log 2 lead provides a single template for parameter representation and ALigN adaptively adjusts the template according to different layers of DNN for producing even better results. We have taken three different DNNs, AlexNet, VGG-16, and Resnet-18, and we quantize them using 8-bit version of our schemes and found minimal loss in accuracy compared to full precision network. For evaluating the efficiency of our technique, we have applied our proposed techniques for image classification and segmentation tasks. We have also presented a multiplier design for efficient multiplication of values represented in our proposed templates. Keywords: Machine learning, deep neural networks, quantization, multipliers, classification, segmentation