Low-power Spike-based Wearable Analytics on RRAM Crossbars

TL;DR

The paper addresses energy-constrained wearable analytics by deploying spike-based SNNs on RRAM crossbars and enables online adaptation via Direct Feedback Alignment (DFA) to cope with hardware non-idealities. It introduces DFA_Sim, a hardware evaluation engine, and uses a Gaussian Process Regression noise model trained on NeuRRAM data to compare DFA against backpropagation, showing substantial improvements in energy, latency, and area with maintained or improved accuracy on HAR tasks. The work demonstrates that layer-parallel gradient updates in DFA reduce error accumulation and hardware overhead, making DFA-based online adaptation viable for edge wearable analytics on non-ideal IMC platforms. Overall, the study highlights a practical path to low-power, spike-based analytics on in-memory RRAM crossbars with real-time adaptation capabilities.

Abstract

This work introduces a spike-based wearable analytics system utilizing Spiking Neural Networks (SNNs) deployed on an In-memory Computing engine based on RRAM crossbars, which are known for their compactness and energy-efficiency. Given the hardware constraints and noise characteristics of the underlying RRAM crossbars, we propose online adaptation of pre-trained SNNs in real-time using Direct Feedback Alignment (DFA) against traditional backpropagation (BP). Direct Feedback Alignment (DFA) learning, that allows layer-parallel gradient computations, acts as a fast, energy & area-efficient method for online adaptation of SNNs on RRAM crossbars, unleashing better algorithmic performance against those adapted using BP. Through extensive simulations using our in-house hardware evaluation engine called DFA_Sim, we find that DFA achieves upto 64.1% lower energy consumption, 10.1% lower area overhead, and a 2.1x reduction in latency compared to BP, while delivering upto 7.55% higher inference accuracy on human activity recognition (HAR) tasks.

Figures (7)

• Figure 1: Pictorial depiction of SNNs used in wearables for temporal data-processing. Pre-trained SNNs in the cloud are adapted online according to the constraints of resource-constrained edge devices.
• Figure 2: Plot showing the advantages of DFA-based online SNN adaptation on an IMC platform over traditional BP.
• Figure 3: (a) Pictorial representation of an SNN. (b) Pictorial representation of an N$\times$N crossbar.
• Figure 4: Pictorial representations of BP (left) and DFA (right) learning.
• Figure 5: Our proposed $DFA\_Sim$ engine. The hierarchical architecture consists of Tiles, Processing Engines (PEs) and RRAM crossbars. For DFA-based adaptation, the crossbars inside the PEs need not to be transposable. For BP-based adaptation the crossbars are transposable, thereby requiring extra periperals. This figure is for representation purpose only; actual number of tiles, PEs and crossbars may differ.