As medical wearables become more widely adopted for at-home/early diagnosis/health surveillance, the volume of data produced by these devices are expected to reach thousands of petabytes/month. Transmitting this large volume of data over the cloud for processing will potentially emerge as a communication bottleneck and increase latency of decisions. Transmitting naively all data generated by a wearable medical device is also costly in terms of power/energy- transmitter is usually the highest consumer of energy in a sensor (at least 10~20x more energy than sensing). Key to addressing this data deluge is to increase capabilities of the wearable devices to process information locally and have on-device inference capabilities, such as through embedding AI capabilities into the wearable device that will allow extraction of key information from the sensor data. There needs to be balance between what can be processed locally on-device with low power/energy and how to optimally decide the volume of data communication from the device (to cloud as an example). The barriers to this approach lie in the computational complexity of AI algorithms that makes it challenging to fit AI models on wearables with limited resources. Some of the answers might lie in going back to early days of signal processing in silicon – developing analog circuit techniques for AI development which will require collaborative innovations in both AI model development and analog circuit design techniques. In this talk, I will present our research on developing analog AI circuits and their demonstrations with patient data with use cases from cardiovascular health monitoring and sepsis onset detection.

The second part of this talk will present machine learning (ML) approaches for enhancing performance of data converters. ML has the potential to emerge as an alternative to current signal processing based complex calibration algorithms for enhancing data converter performance in advanced processes. By learning an efficient representation of the input and data converter behavior, a simple neural network can correct data converter errors arising from multiple sources of non-idealities with similar accuracy as complex calibration algorithms but with a much lower hardware cost. This talk will present machine-learning techniques to improve performance of both analog-to-digital and digital-to-analog converters.

Biography:

Arindam Sanyal is currently an assistant professor in the School of Electrical, Computer and Energy Engineering at Arizona State University. Prior to this, he was an analog design engineer with Silicon Laboratories. He received his Ph.D. in Electrical and Computer Engineering from the University of Texas at Austin, and M.Tech from the Indian Institute of Technology at Kharagpur. Dr. Sanyal’s research expertise includes analog/mixed signal integrated circuits design, bio-medical sensor interface, analog neuromorphic computing, and hardware security. He serves as an Associate Editor in Scientific Reports, Frontiers in Electronics, and Electronics Letters, and as member of Analog Signal Processing Technical Committee (ASP-TC), and VLSI Systems and Applications Technical Committee (VSA-TC) within IEEE Circuits and Systems society, Design Automation Conference (DAC) and VLSI-D.