| KEYWORD |
GR-06 - ELECTRONIC DESIGN AUTOMATION - EDA
Implementation of Reduced-Precision Convolutional Neural Networks
keywords ARTIFICIAL NEURAL NETWORKS, DIGITAL SYSTEM DESIGN, EMBEDDED SYSTEMS
Reference persons DANIELE JAHIER PAGLIARI, MASSIMO PONCINO
Research Groups GR-06 - ELECTRONIC DESIGN AUTOMATION - EDA
Thesis type MASTER THESIS
Description Convolutional Neural Networks (CNNs) used in image classification tasks typically require significant computational and energetic resources, which prevent their usage in embedded devices. One of the most effective hardware solutions for optimizing energy consumption in CNNs consists in replacing floating-point operations with low precision fixed-point equivalents (4-8 bit). In particular, recent studies have demonstrated that using a variable precision, depending on the image to be classified, yields promising results in terms of accuracy. However, the choice of such precision for a given image remains a partially open problem.
The objective of the thesis is the study and implementation of algorithms for the selection of an optimal precision at runtime, by analyzing the input image and the behavior of the neural network itself. The possibility of using different precisions in different sections of the network (layers) will be also considered.
The candidate will evaluate the algorithms using one of the most popular software frameworks for deep learning applications (Tensorflow, PyTorch, etc.) which will be selected at the beginning of the work, based on its available features for the objective of the thesis.
Required skills A basic knowledge of the Python programming language will be useful but is not required.
Deadline 05/05/2019
PROPONI LA TUA CANDIDATURA