Hardware calibrated learning to compensate heterogeneity in analog RRAM-based Spiking Neural Networks

Abstract

Spiking Neural Networks (SNNs) can unleash the full power of analog Resistive Random Access Memories (RRAMs) based circuits for low power signal processing. Their inherent computational sparsity naturally results in energy efficiency benefits. The main challenge implementing robust SNNs is the intrinsic variability (heterogeneity) of both analog CMOS circuits and RRAM technology. In this work, we assessed the performance and variability of RRAM-based neuromorphic circuits that were designed and fabricated using a 130 nm technology node. Based on these results, we propose a Neuromorphic Hardware Calibrated (NHC) SNN, where the learning circuits are calibrated on the measured data. We show that by taking into account the measured heterogeneity characteristics in the off-chip learning phase, the NHC SNN self-corrects its hardware non-idealities and learns to solve benchmark tasks with high accuracy. This work demonstrates how to cope with the heterogeneity of neurons and synapses for increasing classification accuracy in temporal tasks.

The neuromorphic Mosaic: in-memory computing and routing for small-world graphical networks

“…For case (iii), we quantize the weights using a mixed hardware-software experimental methodology whereby memory elements in a Mosaic software model are assigned conductance values programmed into a corresponding memristor in a fabricated array. Programmed conductances are obtained through a closed-loop programming strategy [44][45][46][47] .…”

“…Offline training of neural networks results in full-precision weights that hinder the performance when deployed on RRAM crossbar arrays due to analog-related non-idealities such as programming stochasticity, temporal conductance relaxation, and read noise [44][45][46][47] . To mitigate this detrimental effects at the weight transfer stage, we adapted the noise-resilient training method for RRAM devices 66,67 .…”

28 nm FD-SOI embedded phase change memory exhibiting near-zero drift at 12 K for cryogenic spiking neural networks (SNNs)

“…Examples of energy-e cient unconventional computing solutions are brain-inspired systems, including third-generation spiking neural networks (SNNs) 5,[8][9][10] . As a fundamental building block for such future hardware implementations, emerging non-volatile memories (eNVMs) stand out as promising devices for both compute-in-memory applications and storage-class memory modules 2,4,9,10 . In the case of cryogenic applications, multiple groups have investigated the behaviour and performance of eNVM technologies at cryogenic temperatures, such as ferroelectric 11 , magnetic 12 , and oxide-based resistive memories 13,14 .…”

“…Quantum computers running at cryogenic temperatures 6,7 are another example of potentially revolutionary computing platforms. Examples of energy-e cient unconventional computing solutions are brain-inspired systems, including third-generation spiking neural networks (SNNs) 5,[8][9][10] . As a fundamental building block for such future hardware implementations, emerging non-volatile memories (eNVMs) stand out as promising devices for both compute-in-memory applications and storage-class memory modules 2,4,9,10 .…”