Quantized Context Based LIF Neurons for Recurrent Spiking Neural Networks in 45nm

TL;DR

The paper addresses the challenge of efficient hardware realization for context-aware recurrent spiking neural networks by introducing a quantized Context-Dependent Leaky Integrate and Fire (qCLIF) neuron with dual apical/somatic processing, implemented in 45 nm CMOS. It adopts a hardware-software co-design to enable a scalable RSNN with 10–200 qCLIF neurons and up to 82,000 synapses, using fixed-point inputs and digital weight gating to maintain performance while reducing hardware complexity. Key contributions include the first hardware demonstration of a context-based RSNN with qCLIF, a detailed digital architecture (SWM, AC, MU, SC, TC), and empirical results on the DVS Gesture dataset showing near-full-precision accuracy at 8-bit quantization and favorable energy-area scaling. The work demonstrates the practicality of digital neuromorphic hardware for real-time, context-enabled spiking computation and points to future optimizations in accumulator design and technology node scaling for further gains.

Abstract

In this study, we propose the first hardware implementation of a context-based recurrent spiking neural network (RSNN) emphasizing on integrating dual information streams within the neocortical pyramidal neurons specifically Context- Dependent Leaky Integrate and Fire (CLIF) neuron models, essential element in RSNN. We present a quantized version of the CLIF neuron (qCLIF), developed through a hardware-software codesign approach utilizing the sparse activity of RSNN. Implemented in a 45nm technology node, the qCLIF is compact (900um^2) and achieves a high accuracy of 90% despite 8 bit quantization on DVS gesture classification dataset. Our analysis spans a network configuration from 10 to 200 qCLIF neurons, supporting up to 82k synapses within a 1.86 mm^2 footprint, demonstrating scalability and efficiency

Figures (8)

• Figure 1: Top: Basic idea of using context. Bottom : Comparison of Neuronal Models and Synaptic Processing: Traditional LIF versus proposed qCLIF. The Traditional LIF model exhibits a single-compartment with exponential decay and spike generation upon reaching threshold voltage, with an analog weight multiplication approach. The qCLIF model introduces an additional apical compartment with linear decay dynamics and utilizes a digital weight AND gate mechanism for synaptic processing, offering advantages in speed, reconfigurability, robustness, and scalability of digital hardware methodology.
• Figure 2: A comparison of test accuracy versus parameter count for diverse network architectures addressing the DVS Gesture dataset. Notably, context based Recurrent Spiking Neural Networks (cSNNs) achieve high accuracy with significantly fewer parameters compared to other models referenced in the literature.
• Figure 3: Architecture of a Recurrent Spiking Neural Network layer made of qCLIF Neurons. The inset view highlights a single qCLIF neuron, which processes both input and recurrent spikes using a combination of somatic and apical inputs along with somatic and recurrent, apical weights. It operates with a set of parameters including somatic leak and apical leak, governed by a clock and reset mechanism generating a train of output spikes.
• Figure 4: Digital design of N bit qCLIF neuron. LS : Leakage Subtractor, AA: Apical Accumulator, MU: Multiplication Unit, SLS: Somatic Leakage Subtractor, SA: Somatic Accumulator, TC: Threshold Comparator.
• Figure 5: (a) Setup for Gesture Recognition Using DVS: This setup is utilized for recording gestures, which generate spatio-temporal event streams. The images are adapted from source 11. (b) Simulation Setup: This involves the use of a spatio-temporal stream as somatic input spikes. Context spikes are directed to the apical compartment, prompting the network to recognize a specific class. The output indicates whether the input stream corresponds with the context spikes.