Large Language Models for Time Series: A Survey
Pioneered the first survey on how LLMs can solve time-series tasks through prompting, time-series quantization, alignment of time-series with text, using vision as bridge, and using LLMs as output tools.

ZeroHAR: Contextual Knowledge Augments Zero-Shot Wearable Human Activity Recognition
Innovated a Zero-Shot Wearable Human Activity Recognition framework that integrates contextual knowledge about activities from LLMs and sensory environment information to improve Zero-Shot F1 Score by 20%.

MLO-MAE: Downstream Task Guided Masking Learning in Masked Autoencoders Using Multi-Level Optimization
Developed a Multi-level Optimized Mask Autoencoder that improves image classification by integrating model pretraining with fine-tuning to allow fine-tuning performance guide an optimal masking strategy for pretraining.

PILOT: Physics-Informed Data Denoising for Real-Life Sensing Systems
Established a physics-informed data denoising scheme that integrates physics equations as constraints into training loss, ensuring the cleaned data complies with fundamental physical constraints, thereby improving data quality.

SHARE: Unleashing the Power of Shared Label Structures for Human Activity Recognition
Built an encoder-decoder architecture to model label name semantics that harnesses the shared sub-structure of label names to distill knowledge across classes, improving accuracy of human activity recognition tasks by 1.7%.

MusicLLM: Music Understanding with LLMs
Developed an LLM with music integration that generates text responses, including music genre, instruments used, mood, and theme, based on music files. Used Encodec audio features in conjunction with FLAN-T5 LLM.

STAug: Towards Diverse and Coherent Augmentation for Time-Series Forecasting
Crafted a time-series data augmentation strategy that incorporates both time- and frequency-domain knowledge through mix-up and Empirical Mode Decomposition, respectively, improving time-series forecasting by 10.7%.

PrimeNet: Pre-Training for Irregular Multivariate Time Series
Crafted a self-supervised pretraining method for irregular, asynchronous time-series through sampling density-aware contrastive learning and time-sensitive data reconstruction techniques, improving few-shot downstream performance.

TARNet: Task-Aware Reconstruction for Time-Series Transformer
Developed a self-supervised task-aware data reconstruction technique that uses end-task knowledge to customize the learned representation to boost end task performance, improving time-series classification by 2.7%.

MARS: Multitask Pre-training for Accent Robust Speech Recognition
Built a pre-trained model for accent-robust speech representation that improves performance on several downstream tasks, like Speech Recognition by 20.4% and Speaker Verification by 6.3%, across 12 minority accents with few minutes of data.

ESC-GAN: Extending Spatial Coverage of Physical Sensors
Devised a globally-attentive multiscale Super-Resolution GAN that uses irregularly spaced sparse data to generate high-resolution spatiotemporal data for regions with no physical sensors, reducing reconstruction loss by 3.7%.

UniTS: Short-Time Fourier Inspired Neural Networks for Sensory Time Series Classification
Innovated a Fourier-transform (FT) inspired weight initialization and learnable Short-Time FT layer redesign to integrate time and frequency domain information, improving F1 score for sensory time-series classification by 2.3%.

Automated Ticket Resolution from Semi-Structured Log Data
Developed an ML pipeline to automate ticket resolution by conducting data cleaning, preprocessing, and visualization on time-series semi-structured system-level log corpus, followed by statistical feature extraction and classification.

RIoT: Towards Robust Learning for Internet-of-Things
Built a comprehensive evaluation strategy and a model-agnostic training framework to facilitate robustness of machine learning models to missing feature sets in time-series data, achieving a 22% improvement in robustness.

Model Explainability for Redshift ML
Built a SHAP-based Explainability tool for AWS Redshift ML, enabling users to write SQL queries to introspect model predictions. Improved execution speed by 2x and reduced memory footprint by 90% over existing workflow in AWS Sagemaker.

Generating Rationale in Visual Question Answering
Developed a model that generates a rationale behind selecting a multiple choice answer in visual question answering. Trained a joint visual-linguistic model to build multimodal image- and text-representation and a GPT-2 model to generate text.

Real-Time Principal Component Analysis
Innovated a novel algorithm for dimensionality reduction on streaming data that learns a robust, noise-resilient eigenspace representation using a simple, fast heuristic, achieving Bhattacharyya Distance of over 0.90.

Extracting Real-Time Insights from Stock Market and Twitter
Built and deployed a web application, based on Real Time PCA, to generate real-time analytics on streaming data fetched from Twitter, Google Search Trends, Bangladeshi Online News Portals and Dhaka Stock Exchange.

Stock Price Forecasting from Historical Prices and News Trends
Devised a multi-modal model to forecast stock prices based on historical prices and latest news trends. Used an LSTM to model the stock prices, Word2Vec to represent news headlines and an MLP to concatenate the features.