ETD system

Electronic theses and dissertations repository

 

Tesi etd-06072019-161434


Thesis type
Tesi di laurea magistrale
Author
ROSSI, FEDERICO
URN
etd-06072019-161434
Title
On the Use of Tabulated Posits in Deep Neural Networks for Real-Time Image Classification
Struttura
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
COMPUTER ENGINEERING
Supervisors
relatore Cococcioni, Marco
correlatore Saponara, Sergio
Parole chiave
  • image recognition
  • computer vision
  • neural networks
  • posit
Data inizio appello
19/07/2019;
Consultabilità
Secretata d'ufficio
Data di rilascio
19/07/2089
Riassunto analitico
In last few years, deep neural networks (DNNs) have provided numerous successes inseveral application domains. The main issues that affected this progress have been the huge resource requirements of neural network training.In order to face this problem, a lot of work and research has been done to reduce thisresource burden, thus reducing computational resources and power consumption.One of the most interesting topics in this area is the idea of encoding network weightswith progressively lesser number of bits. Since these weights are typically real numbers,some experiments managed to reduce number of bit from64(double precision) to32(singleprecision) and even to16(half precision), with no drawback on network accuracy.Probably, the most interesting proposed idea as an alternative to float numbers is the Posit,presented by John L. Gustafson as an extension of one of his works about overcoming float number.One of the greatest advantages of this new type is the simplicity of needed circuitry forits implementation, thus obtaining a faster and smaller architecture with regards to the float one.
File