Edge Intelligence with Neuromorphic Computing: From Algorithms to Hardware Design
![Event Image Event Image](/packs/media/images/v2016/featured-event-1-7105b4faed2c0e1c39142be6934e0152.png)
Title: Edge Intelligence with Neuromorphic Computing: From Algorithms to Hardware Design
Abstract: Spiking Neural Networks (SNNs) have emerged as a compelling alternative to deep learning especially for edge computing due to their huge energy efficiency benefits on neuromorphic hardware. In this presentation, I will discuss the challenges and opportunities associated with the design of SNN algorithm and hardware systems, especially in reference to compute-in-memory accelerators. Particularly, I will describe our group's recent works towards enabling and democratizing spike-based machine intelligence design, simulation, and evaluation across different applications.
In the first half, I will talk about the importance of temporal dimension in algorithm design for SNNs which unlock unique behavior such as, robustness and privacy and bring in huge benefits in terms of latency, energy, and accuracy in applications like video segmentation, human activity recognition, and event sensing. In the second half, I will delve into the hardware perspective of SNNs when implemented on compute-in-memory (CiM) and digital systolic array accelerators with our recently proposed SpikeSim and SATA benchmarking tools. It turns out that the multiple timestep computation in SNNs can lead to extra memory overhead and exacerbates the effect of CiM non-idealities that annuls all the compute-sparsity related advantages. I will highlight some techniques such as, input-aware early time-step exit and temporally evolving batch normalization to reduce the overhead. I will discuss an algorithm-hardware co-search methodology that explores the design space of CiM hardware and the neural network topology together, with a search-based optimization to yield best performance-energy efficiency tradeoffs.
Finally, I will present a future vision where our ongoing efforts in neuromorphic algorithm-hardware co-design hold promise for developing novel IoT applications -including, object tracking, drone navigation, prosthetics, human-robot interaction- with high efficiency.