Research Experience

System Design of an Electrochemical Random Access Memory (ECRAM) Based Edge Learning Accelerator

Advisor: Saugata Ghose, Assistant Professor, CS, UIUC

Resource-constrained edge devices increasingly need to learn and adapt from data collected in dynamic environments, but conventional neural network training is too energy- and compute-intensive to run efficiently on these platforms. My work explores how in-memory computing can enable efficient on-device training by reducing data movement and accelerating continual learning workloads. I design an end-to-end edge learning system that integrates ECRAM-based analog acceleration with hardware-software co-design, ISA support, and system-level optimizations to address device non-idealities and training bottlenecks. This system is highly energy efficient compared to an NVIDIA Jetson Orin Nano edge platform.

Continuous Monitoring of Blood Pressure Using Flexible Patch Electronics for Critical/High Risk Patients

Advisor: Muhammad Abdullah Arafat, Assistant Professor, EEE, BUET

In this work, blood pressure has been continuously monitored by our designed smart healthcare system with a wireless and flexible biosensor. This system consists of three main modules - a biosensor to record and amplify the photoplethysmography (PPG) signal by using an IR Reflectance Sensor and an amplifier respectively, a Bluetooth Low Energy Device to transmit the amplified PPG signal wirelessly and a smartphone application for data visualization. We have also incorporated a regression model to correlate the PPG signal, recorded as voltage, to the blood pressure of a person. Moreover, we have opted to integrate the biosensor in a wristband rather than the conventional choice of finger caps. Finally, to enhance the mechanical stretchability and compatibility of the biosensor, we have made the PCB of the biosensor flexible to ensure long-lasting adhesion to the curved skin surface.

An Approach for Analyzing Cognitive Behavior of Autism Spectrum Disorder Using P300 BCI Data

Advisor: Shaikh Anowarul Fattah, Assistant Professor, EEE, BUET

In this study, a machine learning-based scheme is developed as an automated evaluation technique for assessing the cognitive function of Autism Spectrum Disorder (ASD) patients and also to analyze the impact of the visual oddball paradigm as a part of joint-attention training on the cognitive pattern of ASD patients. For this purpose, we have continuously analyzed the EEG signals of ASD patients recorded as the P300 event-related potential (ERP) after each training session. Moreover, the P300 EEG signals are tested in various band-limited conditions with different sets of EEG channels to identify the appropriate method for automated evaluation and finally, after carrying out a meticulous process, we have identified that the selection of alpha, theta and delta frequency bands, EEG channels which ensure visualization of the whole portion of the brain and an ensemble of four machine-learning models like LDA, SVM, MLP and Random Forest provides the best way for automated assessment purpose. Apart from this, we have found an enhancement in the cognitive pattern after the extensive training program for a good number of patients.

Page updated

Google Sites

Report abuse