QuickNN: Python Toolbox for Training and Optimizing ANN for Hardware Implementation

Realizing deep neural networks in hardware is becoming increasingly challenging as applications require ever more layers and weights, incurring computational and storage costs. Current approaches lack the ability to customize neural networks based on hardware design decisions like the number and resolution of neural network inputs, binary/ternary weight quantization, layer quantization, network structure -including number of layers- and activation functions. In this work, we present QuickNN, a Python open-source toolbox for training and optimizing multi-layer perception neural networks designed to aid artificial neural network (ANN) hardware designers. The toolbox allows the customization of most hardware-related design decisions and can provide rapid results to assess the impact of these decisions on performance. In this paper we demonstrate the toolbox on the MNIST dataset, evaluating accuracy for a range of design choices. By way of example, a novel hybrid weight ternary quantization model is implemented in QuickNN that shows improvement in classification performance when compared to state-of-art.

I. Introduction

The massive deployment of deep neural networks (DNNs) in machine learning (ML) applications such as computer vision, speech recognition and natural language processing have ex-ponentially grown in the past few decades [1], [2]. Multilayer Perceptrons (MLPs) shown in Fig. 1(a) are a class of fully connected feed-forward DNNs that gained great popularity in artificial neural network (ANN) hardware accelerators. In most ANN hardware accelerators, the neural network is trained offline and then the weights are mapped to the array of the ANN accelerator to perform inference operations [3], [4]. However, with DNNs getting more complex with emerging AI tasks, the number of layers, weights, computational costs and storage requirements have dramatically increased making the hardware realization very challenging [5].