Tue, 03/30/2021 - 2:00pm
Add to Calendar
You're Invited!
To join, click the flyer below or use this Zoom link.
Meeting ID: 995 1433 2052
Password: 418130
Title: Efficient AI Seminar: Energy-Efficient, Robust and Interpretable Neuromorphic Computing through Algorithm-Hardware Co-Design
Speaker: Dr. Priya Panda, Yale University
Abstract:
Today, deep learning has empowered the digital age enabling ‘intelligence’ in almost all technology that surrounds us. However, the cost associated in terms of computing resources and energy consumption is very high. In that regard, Spiking Neural Networks (SNNs), often referred to as the third generation of neural networks, have the potential of empowering low-power machine intelligence through sparse, event-driven computations. Note, however, today there does not exist good learning and inference techniques that can reap the potential benefits of SNNs.
In this talk, I will present an overview of our recent efforts in training SNNs, addressing the limitations and advantages of spike-driven learning and computations. Specifically, I will talk about a new temporal Batch Normalization Through Time (BNTT) technique that gives competitive performance and reduces the latency (more than 4X faster compared to state-of-the-art SNNs) and energy consumption (9X less compared to standard Aritifical NNs).
Besides enabling accurate and energy-efficient SNNs, there is a need to unveil the ‘black-box’ nature of such networks in order to apply them to ubiquitous scenarios. I will discuss our current work on visualizing spiking networks to understand and interpret the decisions made by the SNN which forays into explainable neuromorphic computing.
Finally, we will understand how hardware can help us facilitate robustness in NN systems. I will talk about an interesting side-effect of non-idealities in memristive crossbars for improving adversarial robustness of mapped deep neural networks (without any additional optimization).
Today, deep learning has empowered the digital age enabling ‘intelligence’ in almost all technology that surrounds us. However, the cost associated in terms of computing resources and energy consumption is very high. In that regard, Spiking Neural Networks (SNNs), often referred to as the third generation of neural networks, have the potential of empowering low-power machine intelligence through sparse, event-driven computations. Note, however, today there does not exist good learning and inference techniques that can reap the potential benefits of SNNs.
In this talk, I will present an overview of our recent efforts in training SNNs, addressing the limitations and advantages of spike-driven learning and computations. Specifically, I will talk about a new temporal Batch Normalization Through Time (BNTT) technique that gives competitive performance and reduces the latency (more than 4X faster compared to state-of-the-art SNNs) and energy consumption (9X less compared to standard Aritifical NNs).
Besides enabling accurate and energy-efficient SNNs, there is a need to unveil the ‘black-box’ nature of such networks in order to apply them to ubiquitous scenarios. I will discuss our current work on visualizing spiking networks to understand and interpret the decisions made by the SNN which forays into explainable neuromorphic computing.
Finally, we will understand how hardware can help us facilitate robustness in NN systems. I will talk about an interesting side-effect of non-idealities in memristive crossbars for improving adversarial robustness of mapped deep neural networks (without any additional optimization).
Bio:
Dr. Priya Panda is an assistant professor in the electrical engineering department at Yale University, USA. She received her PhD degree in Electrical & Computer Engineering from Purdue University, USA in 2019. She obtained her Master’s in Physics and Bache-lor’s in Electrical & Electronics from BITS, Pilani, India in 2013. In the past, she has worked in Intel Labs, Oregon, USA (Summer 2017) where, she developed new algorithms for applying spiking networks on Intel’s Loihi chip. Priya is the recipient of the Amazon Research Award in 2019. Her research interests include- neuromorphic computing and deploying robust and energy efficient ma-chine intelligence.
Series
Dr. Priya Panda is an assistant professor in the electrical engineering department at Yale University, USA. She received her PhD degree in Electrical & Computer Engineering from Purdue University, USA in 2019. She obtained her Master’s in Physics and Bache-lor’s in Electrical & Electronics from BITS, Pilani, India in 2013. In the past, she has worked in Intel Labs, Oregon, USA (Summer 2017) where, she developed new algorithms for applying spiking networks on Intel’s Loihi chip. Priya is the recipient of the Amazon Research Award in 2019. Her research interests include- neuromorphic computing and deploying robust and energy efficient ma-chine intelligence.
Series