Research
Our research focuses on core foundational challenges integrating mathematical tools with real-world objectives to advance the state-of-the-art. We pursue ambitious use-inspired research, targeting frontier perceptual tasks in video, imaging and navigation.
Research Thrusts
-
Advanced Algorithms for Deep Learning
We create fast, provably efficient tools for training neural networks and searching parameter spaces. We develop new theories to rigorously explain successful heuristics.
-
Learning with Dynamic Data
Since datasets are constantly evolving, we research new algorithms and models that can incorporate context and changes at training and test time, including robustness to perturbations.
-
Exploiting Structure in Data
What characteristics of a dataset help with training and inference? We define and uncover rich mathematical structures in datasets to improve downstream modeling and optimization.
-
Optimizing Real-World Objectives
We develop principled methods for automatically satisfying complex constraints and handling interactive feedback from users in real-world situations as is needed for safe robot navigation.
Use Inspired Applications
The foundational research thrusts of IFML all have broad potential impact and feed directly into real-world applications. We selected three use-inspired research areas: video, imaging, and navigation. We work with industrial partners to redesign the whole video pipeline from recognition, compression/decompression to training and model design; we collaborate to improve the imaging pipeline and create novel priors for MRI and circuit quality control; and we develop new methods for autonomous navigation in highly unstructured environments while maintaining safe operation with high confidence.
Publications
-
Knowledge-Augmented Contrastive Learning for Abnormality Classification and Localization in Chest X-rays with Radiomics using a Feedback Loop
Yan Han, Chongyan Chen, Ahmed Tewfik, Benjamin Glicksberg, Ying Ding, Yifan Peng, Zhangyang Wang
arXiv, v5, 2022
-
Goal Blending for Responsive Shared Autonomy in a Navigating Vehicle
Yu-Sian Jiang, Garrett Warnell, and Peter Stone
AAMAS, 2021
-
Temporal-Logic-Based Reward Shaping for Continuing Reinforcement Learning Tasks
Yuqian Jiang, Suda Bharadwaj, Bo Wu, Rishi Shah, Ufuk Topcu, and Peter Stone
AAAI, 2021