Task 004 - Neuromorphic Fabrics
|Event Date:||March 4, 2021|
|Time:||11:00 am (ET) / 8:00am (PT)
Rakshit Pathak, Princeton University Specialized Machine-perception Models and Architectures for Sensor Fusion
Embedded sensors offer distinct advantages for machine perception, including access to modalities unavailable through remote sensing (acoustic, vibration, temperature) and preservation of structure by virtue of invariant association with specific embedded signals. This work aims to harness these advantages scalably and efficiently across very large numbers of embedded sensors, together with remote sensors (e.g., vision), which offer the advantage of rich data acquisition at very low deployment cost. After first reviewing our past work exploring how embedded sensing enhances machine perception through preservation of structure, I will describe our recent work on generalized perception models for fusion of embedded- and remote-sensor data. Such generalized models provide concrete requirements for underlying compute kernels, programmability needs, efficiency needs, and memory-accessing patterns. This motivates our design of hardware-specialized architectures for sensor-fusion-based machine perception.
Rakshit Pathak received his Bachelor of Technology degree in Electronics and Electrical Communication Engineering from Indian Institute of Technology, Kharagpur, India in 2018 and his M.A. degree in Electrical Engineering from Princeton University, New Jersey, in 2020, where he is currently pursuing his Ph.D., guided by Prof. Naveen Verma. His research focuses on algorithm-hardware co-design of machine learning platforms. His research includes design of heterogeneous computing and low-energy in-memory computing systems.