Travailler à TIMA

Propositions de thèses


Test and Reliability of Emerging Memory-based Spiking Neural Networks

Équipe : AMfoRS

Date de début : 01/01/2020

Durée : 36 months

Profil : This PhD thesis is concerned with the following research areas: (i) emerging memory technologies (memristors and/or spintronic devices) used in a non-Von Neumann context, (ii) hardware dependability (robustness, reliability and test) and design-for-dependability, (iii) hardware implementations of bio-inspired neural networks (Spiking Neural Networks). Research hypothesis: the strong restrictions on the size of embedded Spiking Neural Network architectures (limited silicon area and interconnectivity ability) require minimization of the network redundancy which in turn reduces its the intrinsic fault tolerance. There is an acute need to evaluate the reliability and perform manufacturing test of the neuromorphic hardware architectures to guarantee their correct operation and robustness.

See complete information to : follow the link

Personne à contacter : Elena Ioana VATAJELU

 

 

Design and verification of a Spiking Neural Network accelerators with Resistive RAM synapses

Équipe : AMfoRS

Date de début : 01/01/2020

Durée : 36 months

Profil : Spiking Neural Networks (SNN) are seen as a Key building block for strongly improving the energy efficiency of current AI applications and opening up new possibilities (in terms of unsupervised learning, recurrent networks, probabilistic inference, etc.). In that respect, one of the key scientific challenges is to design a scalable and flexible SNN architecture: that architecture must be adaptable to different algorithms, to handle tasks such as inference and learning ones (online, supervised, unsupervised, probabilistic, etc.).
Neural networks being memory-intensive architectures, it is mandatory to implement all of the memory on die, in order to reach state-of-the-art energy efficiencies. This is why circuits will be designed and fabricated in hybrid nanoscale CMOS and Resistive RAM technology, enabling very high synaptic density.
The obtained circuits will be employed in embedded applications, in the industrial, health and automotive sectors.

See complete information to : follow the link

Personne à contacter : Alexandre VALENTIAN (CEA) / Lorena ANGHEL (TIMA Lab) / Elena Ioana VATAJELU (TIMA Lab)

 

 

Spiking Neural Networks On Line Learning Strategies

Équipe : AMfoRS

Date de début : 01/01/2020

Durée : 36 months

Profil : The main issues considered in SNNs implementations is the power consumption reduction, that will be achieved by the right combination of learning algorithms, large scale power efficient accelerators developed in a Non-von Neuman style and the utilization of emerging technologies. Therefore, research communities have to jointly consider circuit robust design, architecture constraints and simplified, hardware-oriented learning algorithms. In fact, the real switch from classic Deep Neural Network to Spiking Neural Network would really be reached when other properties of bio-inspired neural networks such as unsupervised and distributed learning could be efficiently implemented. In that respect, one of the key scientific challenges is to design a scalable and flexible SNN architecture that can adapt to different learning algorithms, to handle not only inference tasks but also the learning ones (online, supervised, unsupervised, probabilistic, etc.).

See complete information to : follow the link

Personne à contacter : Elena Ioana VATAJELU (TIMA Lab) / Alexandre VALENTIAN (CEA) / Lorena ANGHEL (TIMA Lab)