UNIVERSITY OF IDAHODEPARTMENT OF MATHEMATICS
|
---|
Recently, U. S. Office of Science and
Technology announced a grand challenge that called to
"Create a new type of computer that can proactively
interpret and learn from data, solve unfamiliar problems
using what it has learned, and operate with the energy
efficiency of the human brain." The current explosion in
deep-learning applications is expected to hit a dual
wall-(1) plateauing in CMOS scaling, and (2) limits set
for energy consumption in a data center. This calls for
setting aside the digital von Neumann paradigms for
computing and look at analog alternatives by deriving
inspiration from the human brain. Spiking neural networks
have made steady progress to provide alternative low-power
approaches to deep-learning; however, the gains derived
from deploying them on digital platforms are not
substantial. Large-scale integration of nanoscale emerging devices, such as the resistive RAM (or RRAM) devices in large crosspoint array format, with CMOS, can enable a new generation of Neuromorphic Computers that can solve a wide range of machine learning problems. Such mixed-signal Neuromorphic architectures will result in several orders of magnitude reduction in energy consumption. Several studies, including by the presenter, have been performed using RRAMs as non-volatile "analog" synapses in spiking neural network implementations. However, the progress in this area has been impeded from the realization that the actual RRAM devices fall well short of the expected behavior from the idealized behavior. Based upon the recent results, the presenter will discuss the challenges associated with these devices, their emulation using compact CMOS circuits; architectures that bridge hybrid CMOS-RRAM circuits with spiking neural network algorithms, and their use in neuromorphic system-on-a-chip that can implement large scale deep-learning. Further the presenter will introduce proposed pathways to overcome these challenges to realize Neuromorphic System-on-a-chip (NeuSoC) and extend them to large networks using CMOS photonic interconnects.
|