Direct interfacing of bio-inspired NEMS Sensors to bio-inspired RRAM spikingnetworkRKS
Published : 5 April 2019
Extracting useful and compact information from sensor data is key for future mobile and Internet of Things (IoT) applications. Mining data from raw sensors remains an open problem, so that systems capable of handling large volumes of noisy and incomplete real-life data are required. Today, the most promising approach is deep learning. Despite its benefit, the adoption of deep learning within IoT faces significant barriers due to the constraints imposed by mobile devises (memory, power consumption, and limited transmission range).
One possible approach to tackle these challenges is to rethink and reorganize computer architecture taking inspiration of living organisms. Insects are not able to perform calculations like digital systems but excel in controlling small and agile motor systems based on the fusion of data sparse sensory inputs. Moreover, they operate under severe constrains, of energy conservation and limited communication range, among others. Therefore, they provide highly interesting model systems for neuromorphic embedded computation. Resistive RAM (RRAM) are non-volatile memory elements whose values/conductances change as a function of the applied pulses. Thanks to these properties they are prime candidates for implementing plastic synapses in neuromorphic systems. Arrays of micromechanical pillars mimicking the cricket hairs have been demonstrated to be excellent air flow sensors. The main objective of the project is to develop a bio-inspired RRAM-based spiking neural network directly interfaced with a bio-inspired MEMS sensor for readout and local information processing.
The main research objective is the design, fabrication and test of a RRAM-based spiking neural network for the readout of an already available nanomechanical resonator array. The alleged advantages of the proposed bio-inspired design throughout the whole system will be demonstrated by simulations calibrated on the experimental results.