ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-07272021-093922


Tipo di tesi
Tesi di laurea magistrale
Autore
PAOLINI, EMILIO
URN
etd-07272021-093922
Titolo
Development of a Fixed-Point Deep Neural Networks Library in C++ and its use to validate Photonic Neuromorphic Accelerators
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Relatori
relatore Prof. Cococcioni, Marco
relatore Ing. Andriolli, Nicola
relatore Ing. De Marinis, Lorenzo
Parole chiave
  • photonics
  • neural networks
  • deep learning
  • artificial intelligence
Data inizio appello
24/09/2021
Consultabilità
Tesi non consultabile
Riassunto
In recent years, deep neural networks (DNN) have become one of the most powerful tools in machine learning, achieving unprecedented milestones in various fields such as computer vision, genomic interpretation, robotics and autonomous driving. However, the energy consumption and footprint for computation and data movement in DNN is now becoming a major limiting factor impacting DNN scalability.
Regarding computation, the energy consumption is dominated by multiply accumulate operations, which constitute the linear part of DNN computations. In this context, photonic solutions are being investigated as an energy-efficient alternative to electronics-based DNNs because of the inherent parallelism, the high processing rate with low latency, and the possibility to exploit passive optical elements.
However, the drawback of this approach is the reduced precision that can be achieved by such analog photonic engines. In this context, reduced-precision goes beyond classical 16-bit half-precision floats up to very small values, i.e., <=8 bits or even 2 bits. Recent breakthrough in the field demonstrated a nearly negligible degradation on DNN accuracy when using these novel precision-scalable architectures, where the bit resolution can be adjusted to trade off the neural network inference accuracy with speed and power consumption.
The aim of this thesis is to develop a C++ library for DNNs using a low-precision fixed-point format and then to develop a model to reduce the gap between the software implementation on electronic computers and the photonic accelerators that we are currently able to implement in hardware.
File