Exploring Liquid Neural Networks on Loihi-2

TL;DR

Liquid Neural Networks (LNNs) are continuous-time, dynamic architectures that adapt to temporal data, offering energy-efficient computation on neuromorphic hardware. This paper analyzes theoretical foundations of LTCNs and practical pathways for mapping LNNs to neuromorphic chips, including a Neural Circuit Policy (NCP) framework for spatial feature extraction and decision making. It reports a CIFAR-10 classification result on Intel Loihi-2 achieving $91.3\%$ accuracy with $213\,\mu\mathrm{J}$ per frame, establishing a new benchmark for efficiency and accuracy. The study also discusses hardware challenges, co-design considerations, and future directions towards scalable, brain-like AI on low-power neuromorphic platforms.

Abstract

This study investigates the realm of liquid neural networks (LNNs) and their deployment on neuromorphic hardware platforms. It provides an in-depth analysis of Liquid State Machines (LSMs) and explores the adaptation of LNN architectures to neuromorphic systems, highlighting the theoretical foundations and practical applications. We introduce a pioneering approach to image classification on the CIFAR-10 dataset by implementing Liquid Neural Networks (LNNs) on state-of-the-art neuromorphic hardware platforms. Our Loihi-2 ASIC-based architecture demonstrates exceptional performance, achieving a remarkable accuracy of 91.3% while consuming only 213 microJoules per frame. These results underscore the substantial potential of LNNs for advancing neuromorphic computing and establish a new benchmark for the field in terms of both efficiency and accuracy.

Figures (3)

• Figure 1: Schematic Diagram of Liquid Neural Networks (LNNs). Figure illustrates simplified architecture of an LNN, starting with the input layer receiving time-series data or as in our case images from datasets like CIFAR-10. The data then flows into the liquid layer, where dynamic, non-linear processing occurs through a complex network of interconnected neurons. Finally, the processed information is relayed to the output layer.
• Figure 2: LNNs Hardware Implementation. Figure illustrates the data processing flow of implementing an LNN into neuromorphic hardware, beginning with the CIFAR-10 dataset at the bottom which we deployed, which undergoes feature extraction via convolutional layers. The extracted features are then integrated within the Neural Circuit Policy (NCP) Framework, a decision-making system that steers the data through the liquid layer illustrated in the top-right. After training, the LNN is implemented on neuromorphic hardware, symbolized by the chip icon.
• Figure 3: Block Diagram of Implementation.