|MALAVENA GERARDO||Cycle: XXXIII |
Tutor: SOTTOCORNOLA SPINELLI ALESSANDRO
Advisor: MONZIO COMPAGNONI CHRISTIAN Major Research topic
:Investigation of Neuromorphic Computing Systems based on NOR Flash memory ArraysAbstract:
Artiﬁcial Neural Networks (ANNs) are often successfully applied to problems for which no simple algorithmic solution exists, but can be easily addressed by some inference from a set of examples. Thanks to a learning phase, an ANN is able to adjust its internal parameters to behave in a certain way when inputs with similar features are supplied. Hence, taking inspiration from the concepts exploited by human brain to elaborate information, ANNs provide new approaches to deal with a great variety of applications, from ﬁnancial analysis to data mining.
Software-based ANNs oﬀer the chance to fully exploit the fault tolerance and the parallel computing capability provided by these bio-inspired systems. Nevertheless, modern digital computers are not speciﬁcally designed to implement such features, making ANNs not really eﬃcient. Recent improvements on CMOS technology allowed to design hardware-based ANNs able to outperform the software counterpart in terms of speed and power consumption, by distributing the input elaboration across simple elements instead of using a central processing unit. Notwithstanding, the memory accesses performed to retrieve the weight that each element has to assign at its input limit the actual speed and power saving achievable. Nowadays, many solutions to overcome these issues are under investigation but most of them are still in their infancy. Among the more promising alternatives, the non-volatile memories (NVMs) appear suitable for massively-parallel and energy-eﬃcient neuromorphic computing systems. In particular, the NOR Flash arrays are highly reliable devices allowing to achieve very large scale designs without any size constraint, an intriguing feature to augment the number of synapses that the system is able to emulate.The aim of this thesis is to investigate the performance of a neuromorphic computing
system based on NOR Flash arrays.