Collaborative Research: FET: Neuro-AI Foundations for Long-Range Dependency Tasks

The massive explosion of foundation models in the past few years has led to widespread adoption of AI in a plethora of applications, having ubiquitous presence in our everyday lives. However, the computational resource requirements to enable these platforms are also growing at an unsustainable rate. Coupled with this computational bottleneck, lies the increasing focus to enable long context handling of these foundation models, further exacerbating the issue. The situation becomes especially challenging for resource constrained edge environments, like Unmanned Aerial Vehicles, robots, wearables and mobile phones, where on-chip inference may be desired due to latency requirements, bandwidth limitations and privacy. This project aims to address fundamental challenges in enabling Spiking Neural Networks (SSNs) to efficiently handle long-range dependencies and generative tasks for edge-based AI systems by rethinking SNNs as underlying State Space Models. The research has far-reaching impacts on the semiconductor and AI industry, where neuroAI and neuromorphic computing can potentially achieve significant improvements in data-intensive ML workloads that conventional approaches may not be able to implement efficiently in resource-constrained environments. The cross-layer nature of the project ranging from device design, circuit and system explorations and machine learning will serve as an ideal interdisciplinary platform to train the next generation of researchers and engineers. The cross-cutting research agenda combines insights from devices and circuits to machine learning and dynamical systems. Reciprocal information flow between the following thrust areas will be pursued -- (1) Thrust 1: algorithmic innovations to enable neuro-AI models for long range dependency tasks offering orders of magnitude power/energy efficiency along with advantages like adversarial robustness; (2) Thrust 2: co-design of the hardware components that can natively implement the algorithmic innovations with emphasis on scalability, reconfigurability and energy efficiency to enable efficient hardware implementation of the developed neuro-AI models; and (3) Thrust 3: System-level optimizations that address tradeoff between accuracy, throughput versus circuit and device complexities that can lead to broad practical deployments. If successful, the research has the potential to drive transformative advancements in energy-efficient AI platforms performing autonomous reasoning and decision-making. By incorporating insights from dynamical systems, hardware platforms and application needs, this work will pioneer energy-efficient frameworks capable of token-by-token online processing and bridge critical gaps in SNN applicability for real-time, complex sequential and generative tasks. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.