Projects

    1. Conditional Synthetic Data Generation

    At the onset of a pandemic, the data corresponding to the new disease is scarce. At the same time, ML algorithms require large amounts of data to rapidly adapt to the new disease. This scenario is common in a number of application domains, where specific data are scarce, or the dataset is class-imbalanced. We designed a convolutional neural network based framework consisting of a feature extractor and a conditional generative flow working in tandem to learn the probabilistic distribution of input data, and generate conditional synthetic samples. A classifier trained on our generated synthetic data achieved at par accuracy (96.3%) with that trained on real data (96.46%) for chest CT scans. We also design conditional synthetic data across multiple domains, where labeled data in one or more domains are unavailable.

    Publications:
    • Conditional Synthetic Data Generation for Robust Machine Learning Applications with Limited Pandemic Data [PDF][BibTex]
      Hari Prasanna Das, Ryan Tran, Japjot Singh, Xiangyu Yue, Geoffrey Tison, Alberto Sangiovanni-Vincentelli Costas J. Spanos
      Proceedings of the AAAI Conference on Artificial Intelligence 2022


    • CDCGen: Cross-Domain Conditional Generation via Normalizing Flows and Adversarial Training [PDF][BibTex]
      Hari Prasanna Das, Ryan Tran, Japjot Singh, Yu-Wen Lin and Costas J. Spanos
      Workshop on Machine Learning for Data: Automated Creation, Privacy, Bias, International Conference on Machine Learning (ICML) 2021
      Workshop on Data-Efficient Machine Learning (DeMaL), Conference on Knowledge Discovery and Data Mining (KDD) 2021


    2. Transfer Learning for Smart Building Applications

    Often, machine learning algorithms developed for a domain do not generalize well to another domain. Also, in many domains, labeled data is unavailable and/or resource-intensive to collect. This challenge gets amplified in Smart Buildings because of the diversity in built environments, occupants, and building types. We developed transfer learning methods using adversarial domain adaptation for various source and target domain cases. For the opportunistic scenario when a handful of labeled samples are available in the target domain, we designed semi-supervised learning based domain adaptation methods. Our applications in Smart Buildings include enabling transfer learning of personal thermal comfort prediction model from one occupant to another, and WiFi-based occupant gesture recognition model transfer across spatial environments.

    Publications:
    • Unsupervised Personal Thermal Comfort Prediction via Adversarial Domain Adaptation [PDF][BibTex]
      Hari Prasanna Das, Stefano Schiavon, Costas J. Spanos
      In Proceedings of the 8th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys) 2021 (Best Poster Award)


    • Multi-source Few-shot Domain Adaptation [PDF][BibTex]
      Xiangyu Yue, Zangwei Zheng, Colorado Reed, Hari Prasanna Das, Kurt Keutzer, Alberto Sangiovanni-Vincentelli
      arXiv preprint arXiv:2109.12391, 2021


    • Consensus Adversarial Domain Adaptation [PDF][BibTex]
      Han Zou, Yuxun Zhou, Jianfei Yang, Huihan Liu, Hari Prasanna Das and Costas J. Spanos
      Proceedings of the AAAI Conference on Artificial Intelligence 2019


    3. Personal Thermal Comfort Models and Time-Series based Prediction

    Humans spend more than 90% of their day indoors, where their well-being, performance and energy consumption are linked to thermal comfort. But, study shows that only 40% of commercial building occupants are satisfied with their thermal environment. In this project, we conducted an experiment to collect time-stamped physiological signals for subjects using wearable sensors, along with their thermal preferences and other environmental parameters. We designed thermal comfort prediction models using classical, and time-series based ML methods. We also proposed data transformation methods to make tabular data (commonly found in smart buildings) compatible for use with machine learning (especially neural network) based methods.

    Publications:
    • Improved Dequantization and Normalization Methods for Tabular Data Pre-Processing in Smart Buildings
      Hari Prasanna Das, and Costas J. Spanos
      In Proceedings of the 9th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys) 2022


    • Time Series-based Deep Learning model for Personal Thermal Comfort Prediction [PDF][BibTex]
      Aniruddh Chennapragada, Divya Periyakoil, Hari Prasanna Das, and Costas J. Spanos
      In Proceedings of the 13th ACM International Conference on Future Energy Systems (ACM e-Energy) 2022


    • Personal thermal comfort models with wearable sensors [PDF][Dataset][BibTex]
      Shichao Liu, Stefano Schiavon, Hari Prasanna Das, Ming Jin and Costas J. Spanos
      Building and Environment (2019), doi:10.1016/j.buildenv.2019.106281

    4. Machine Learning Empowered Occupancy Sensing and Activity Detection

    To achieve the criteria of Data Privacy and Robustness in Smart Buildings, we formulated computer vision based algorithms to detect human activity and occupancy in a building using channel state information data from WiFi modules (as they are ubiquitous, and provide sufficient privacy) and proposed a sensor fusion approach to combine data from WiFi and camera modules for robust human activity detection.

    Publications:
    • Machine Learning empowered Occupancy Sensing for Smart Buildings [PDF][BibTex]
      Han Zou, Hari Prasanna Das, Jianfei Yang, Yuxun Zhou and Costas J. Spanos
      Climate Change + AI Workshop, International Conference on Machine Learning (ICML) 2019


    • WiFi and Vision Multimodal Learning for Accurate and Robust Device-Free Human Activity Recognition [PDF][BibTex]
      Han Zou, Jianfei Yang, Hari Prasanna Das, Huihan Liu, Yuxun Zhou and Costas J. Spanos
      Proceedings of the Multimodal Learning and Applications (MULA) Workshop, Conference on Computer Vision and Pattern Recognition (CVPR) 2019


    5. Graphical Lasso based Segmentation Analysis for Energy Game-Theoretic Frameworks

    Occupants of a building typically lack the independent motivation necessary to contribute to and play a key role in the control of smart building infrastructure. Energy game-theoretic frameworks create a friendly competition between occupants/users, motivating them to individually consider their own energy usage and seek to improve it to have a better score/achieve a lucrative incentive in the game. We proposed a graphical lasso based hybrid and explainable clustering approach for designing better incentives for future games.

    Publications:
    • A Novel Graphical Lasso Based Approach Towards Segmentation Analysis in Energy Game-Theoretic Frameworks [PDF][BibTex]
      Hari Prasanna Das, Ioannis C. Konstantakopoulos, Aummul Baneen Manasawala, Tanya Veeravalli, Huihan Liu and Costas J. Spanos
      18th IEEE International Conference On Machine Learning And Applications (ICMLA), Boca Raton, FL, USA, 2019, pp. 1702-1709


    • Do Occupants in a Building exhibit patterns in Energy Consumption? Analyzing Clusters in Energy Social Games [PDF][Slides]
      Hari Prasanna Das, Ioannis C. Konstantakopoulos, Aummul Baneen Manasawala, Tanya Veeravalli, Huihan Liu and Costas J. Spanos
      Workshop on Tackling Climate Change with Machine Learning, Conference on Neural Information Processing Systems (NeurIPS) 2020


    • Design, Benchmarking and Graphical Lasso based Explainability Analysis of an Energy Game-Theoretic Framework [PDF][Poster]
      Hari Prasanna Das, Ioannis C. Konstantakopoulos, Aummul Baneen Manasawala, Tanya Veeravalli, Huihan Liu and Costas J. Spanos
      Workshop on Tackling Climate Change with Machine Learning, Conference on Neural Information Processing Systems (NeurIPS) 2019