Advancing Diffusion Models: Alias-Free Resampling and Enhanced Rotational Equivariance
Developed an enhanced diffusion model by integrating alias-free resampling techniques inspired by StyleGAN3, resulting in improved rotational equivariance and stable training. The model demonstrates superior performance across MNIST, CIFAR-10, and MNIST-M datasets without introducing new trainable parameters, maintaining simplicity and efficiency in architecture.
Tech Stack: Pytorch

LiPCoT: Linear Predictive Coding based Tokenizer for Self-supervised Learning of Time Series Data via Language Models
Designed a novel tokenizer that converts time series data of brain signals into a fixed vocabulary via stochastic modeling and utilized BERT for pre-training and fine-tuning for detecting neurodegenerative diseases tasks.
Tech Stack: Pytorch

HugTokenCraft: Python library to modify BERT Tokenizer
Developed a Python library to simplify the process of vocabulary modification (add/remove tokens from dictionary) of a pre-trained tokenizer (BertTokenizer) used in BERT
Tech Stack: BERT

Parkinson's Disease Classification via EEG: All You Need is a Single Convolutional Layer
Developed LightCNN, a single-layer CNN for Parkinson's disease classification using EEG data. LightCNN outperformed complex models in accuracy, precision, and recall while remaining lightweight and interpretable, making it ideal for deployment in resource-constrained settings, capturing clinically relevant neurophysiological patterns associated with PD.
Tech Stack: Pytorch, EEG data

Real-time Classification of Time Series
Developed supervised models using Linear Discriminant Analysis for binary classification of real-time incoming time series data of brain activity using frequency-domain features for adaptive brain stimulation.
Tech Stack: Python

Efficient Feature Engineering for Time Series Classifier
Developed manual features using signal processing (spectral power) from LFP brain activity time series (RCS brain-implant) for computationally efficient sleep stage classification and achieved over 90% accuracy with lightweight ML models (SVM), matching DL models (CNN, Transformer, RNN) with automatic feature engineering (published in Nature Communications)
Tech Stack: Matlab, R

TurboLPC: Python library for Signal Processing
Developed an efficient Python library for stochas￾tic modeling of time series data for speech and audio analysis, data compression, and feature extraction. TurboLPC is 1,000x faster and comes with an advanced variation of frequency-warping for non-uniform frequency resolution
Tech Stack: Python

Generalized Multi-Class Supervised Models for Time Series
Created multi-class supervised models for time series data of brain activity, achieving an average accuracy of 85%. Implemented extensive feature extraction, including 74 features (18 frequency-domain, 10 time-domain and additional coherence features). Evaluated 8 machine learning models, including ridge regression, SVM, KNN, decision tree, random forest, XGBoost, LightGBM, and artificial neural network.
Tech Stack: Python

Novel Feature Extraction Method for Time Series Classification
Designed a computationally efficient method to obtain features (via stochastic modeling) and ML algorithm for supervised classification of time series data improving both performance (+13% accuracy) and computational cost (5× faster).
Tech Stack: Matlab

Regression Model for Time Series Data
Designed ML algorithm that utilizes stochastic characteristics of time series data as features to estimate observed measurements.
Tech Stack: Matlab

Source Estimation from Time Series
Utilized single/multi-channel sensor time series data to estimate the characteristics of the source for localization and classification tasks.
Tech Stack: Matlab

Statistical Evaluation Framework
Developed complex statistical models and tests (Linear Mixed effects/t-test/MANOVA/Hierarchical Cluster Analysis) for time series data and features (via cross-correlation, spectral coherence, spectrogram, and wavelet analyses) to evaluate the stability of data-driven clustering and discover neurophysiological changes in brain networks.
Tech Stack: R

Automated Artifact Detection & Removal
Developed algorithms to detect and rectify abnormal data and noise artifacts in time series data of brain activities using signal processing tools (template matching, cross-correlation, Spectral power threshold, Kalman Filtering) and unsupervised ML clustering (Multivariate Auto-regressive HMM).
Tech Stack: Matlab

Data Synchronization among Wearable Devices
Designed signal processing algorithms for synchronizing time series data from multiple wearable and implanted devices while accounting for missing data, disconnections, mismatch of timestamps, and non-uniform sampling rates.
Tech Stack: Matlab

Data Processing Pipeline for Wearables
Developed a data collection and processing pipeline for wearable devices for building ML models and conducting statistical analyses for automatic data collection from servers (via RESTful APIs), data quality checks, resampling, artifact rejection, manual data corrections, and pre-processing (filtering).
Tech Stack: Matlab

Data Processing Tool for EEG
Developed pipeline for processing raw time series data (EEG) for building ML models and feature engineering.
Tech Stack: Matlab

Motion Segmentation using Subspace Clustering
Utilized ML algorithms (Mixture of Probabilistic Principal Component Analysis, k-Subspaces and Sparse Subspace clustering) to recognize moving objects in video.
Tech Stack: Matlab

Cloud-based AI-Assisted Education Platform
Developed a smartphone app (Android) and server-side web application for an AI-assisted education platform, supporting 10,000+ users. The app, deployed via Ruby on Rails with a complex PostgreSQL database and hosted using open-source PaaS (CapRover).
Tech Stack: Ruby, Rails, Postgres, Java, Docker

A Personalized Brain Pacemaker for Parkinson’s
News Article
August 20, 2024
The New York Times
link

Taming Parkinson’s Disease with Intelligent Brain Pacemakers
News Article
August 19, 2024
Science Daily
link

Taming Parkinson’s Disease with Intelligent Brain Pacemakers
News Article
August 19, 2024
UCSF News
link

ECE research could enhance Parkinson’s disease treatment in rural settings
News Article
January 4, 2024
College of Engineering, The University of Iowa
link

University of Iowa researchers develop algorithm to detect Parkinson’s Disease
News Article
September 9, 2020
The Daily Iowan
link

UI engineers and neurologists develop a highly efficient algorithm that can detect Parkinson’s Disease through EEG data
News Article
August 31, 2020
College of Engineering, The University of Iowa
link