Robust Spatial-Temporal Incident Prediction
I am interested in understanding when and where spatial-temporal incidents like accidents, medical calls, and crimes like wildlife poaching and illegal fishing occur. There are several fundamental technical challenges associated with such problems. First, such incidents are extremely sparse, so traditional machine learning systems cannot be used for forecasting. Second, incidents must be predicted (or detected) based on heterogeneous data sources including crowdsourced data. Finally, existing prediction methodologies fail to tackle strategic manipulation – agents that commit crimes like poaching can observe patrols and change their behavior. I work to design machine learning pipelines that can learn robust spatio-temporal incident prediction and detection models from sparse and heterogeneous data.
To know more about how to design robust forecasting approaches against strategic shifts, read our latest paper from UAI 2020 here. To see how spatial-temporal forecasting models can be made by learning from sparse data, see our paper from IEEE SmartComp 2021 here. To know more about how crowdsourced data can detect emergency incidents before they are reported, see our paper from ICDM 2021.
Decision Making under Uncertainty for Smart Cities
Multi-agent systems are a crucial part of smart cities. In various situations like emergency response, resource allocation under uncertainty plays a crucial role. First-responders are constrained by limited resources, and must attend to different types of incidents like traffic accidents, fires, crime, and distress calls. I work with first responders to design approaches to tackle decision-making for multiple agents under uncertainty. Our work has been showcased at multiple global smart city summits, covered in the government technology magazine and won the best paper award at ICLR’s AI for Social Good Workshop. Read our survey paper summarizing the research in this space in the last few decades. We have worked to create online approaches to spatial-temporal learning, tractable algorithms for learning in semi-Markovian decision processes, decentralized approaches to optimize multi-agent systems in limited communication settings, and hierarchical learning to design tractable Monte-Carlo tree search. For a full list of publications, see here.
Wildifire Spread Modeling
I am currently working to understand how wildfires spread by using large-scale satelitte imagery. Our goals are three-fold in this project. First, we are creating WildfireDB, the first open-source and comprehensive database that links wildfire occurrence to features extracted from satellite imagery (>17 million data points). Read our paper from the NeurIPS AI for Earth Sciences workshop to know more about this open-source database. Processing large scale geospatial data is often challenging, especially because combining large vector and raster data can be difficult. We have created a novel way of processing large geospatial data. See our paper from ICDE 2021 (to appear) to know more. Second, we are trying to understand how wildfires spread as a function of relevant determinants like vegetation type and weather. Third, we are trying to understand how to best deploy resources to suppress wildfires when the true state of the wildfires cannot be observed. Read our workshop paper here to know more about this project.
Transit Optimization for Smart Cities
As cities grow, modes of transit diversify. I am currently studying fairness in transit allocation and design and trying to understand how transportation demand can be forecasted and optimized. Currently, we are trying to optimize para-transit routing to meet real-time demand, and understand if occupancy forecasting can be used to aid social distancing in public transportation.
Monarch Butterfly Migration
Every winter, thousands of monarch butterflies from Canada and United States migrate to a small group of forests in Mexico. Monarchs are increasingly being threatened by deforestation and climate change. I am interested in animal migrations in general, and currently working on designing approaches to automate the counting of monarch butterflies in clusters. We are supported by a Microsoft AI for Earth grant for this project. See a preprint of our paper on using deep learning to count monarch butterflies in dense clusters here.