| KEYWORD |
Event-based Graph Neural networks for silicon retinas: efficient HW implementation
keywords GRAPH NEURAL NETWORKS, HW IMPLEMENTATION OF MACHINE LEARNING ALGORITHMS, SPIKING NEURAL NETWORKS
Reference persons LUCIANO LAVAGNO
External reference persons Fabrizio Ottati, Usman Jamal, Filippo Minnella
Research Groups Microelectronics
Thesis type RICERCA
Description State of the art AI algorithms devised for silicon retinas work by transforming events into dense representations processed using standard convolutional DNNs, discarding the sparsity and high temporal resolution of events and leading to high computational burden and latency. For this reason, recent works have adopted Graph Neural Networks (GNNs), which process events as “static” spatio-temporal graphs, which are inherently ”sparse”.
Asynchronous, Event-based Graph Neural Networks (AEGNNs), a novel event-processing paradigm that generalizes GNNs to process events as
“evolving” spatio-temporal graphs, use update rules that restrict recomputation of network activations only to the nodes affected by each new
event, reducing both computation and latency for event by-event processing.
In this thesis, the high sparsity and reusability of AEGNNs computations should be exploited in an efficient hardware implementation.
Required skills Digital hardware design.
Machine learning knowledge is useful but not required.
Deadline 20/12/2023
PROPONI LA TUA CANDIDATURA