AI for Smart City Transportation

• Unsupervised Summarization and Descriptive Modelling of City Traffic Data: Developed machine learning algorithms focused on modelling spatio-temporal dynamics in traffic flow. Different use cases for these predictive/descriptive models include detecting unusual time periods, summarizing mobility patterns, modelling contextual changes etc

• Real-time System for Traffic Incident/Interruption Detection: Developed a real time system to detect traffic interruptions (anomaly) in using ground sensor data. Developed ML pipeline with non linear dimensionality reduction followed by deep learning based classifier to achieve an accuracy of 90% outperforming several other existing methods

• Deep Learning based Digital Twin for City Traffic Network: InterTwin: Developed a deep learning based framework that can emulate physics based microscopic simulations of a traffic network. Developed novel architectures using graph convolutions, attention heads, recurrent layers. The proposed models are 5 orders of magnitude faster (compared to microscopic simulations) and is used for extensive parameter exploration for signal timing optimization of a city network.

• BigSUMO: End-to-End Parallel Deep Learning System for Traffic Simulation Applications: Developed an end-end framework (data generation - model training - hyperparameter optimization). Leverages distributed computing to generate, process terabyte-scale heterogeneous traffic simulation data. Equipped with SOTA open source libraries - OpenMPI (parallelizing), HDF5 (storage), PyTorch(distributed training), MLFLOW (experiment tracking), Optuna (hyper-parameter optimization). Tested running on more than 5000 cores on UF's supercomputer.

• Deep Reinforcement Learning for Adaptive Traffic Signal Control: Currently working on developing multi agent deep reinforcement learning framework for controlling signal timing at each intersection considering local and global state of traffic network.

Hybrid Learning Tecnniques for Scientific Data Compression

• Data compression pipelines for exascale scientific applications that produce extremely large amounts of data. Currently developing hybrid learning-based compression techniques that combines ideas from tensor decompositions, auto encoders, generative modelling, product quantizers, and entropy coding.

• Developed models with error guarantees on primary data as well as derived quantities of interest. Tested on petabyte scale atmospheric reanalysis data and has shown guarantees on cyclone detection and tracking.

• Software engineering lead for ROAMM, a customizable framework for real time online assessment and mobility monitoring. Supports real time data collection, capturing health events with secure wearable technology. Highly scalable cloud based framework, supports customization for different research needs. Currently used by more than 10 different research groups for their clinical studies.

• Developed highly scalable micro service based architecture for remote data collection, campaign management in real time.

• Developing ML models using accelerometry data for different use cases like activity recognition, research on behaviors, health and mobility patterns of the individuals

• Working in inter-disciplinary team consisting of experts from Computer Science, Geriatric, Clinical, Biomedical and Cognitive researchers