Publications

Data-driven simulation of LiDAR and event cameras for policy learning and zero shot deployment.

Data-driven simulation capable of inpainting novel multi-agent into the view of the ego agent for transferable policy learning.

A closed-form solution for a class of ODE-based neural models that achieves a 150x speedup and superior accuracy.

Demonstrate that continuous-time neural models are dynamic causal models capable of interpretable, causal and robust visual navigation

For continuous-time normalizing flows, pruning is all you need!

Fast & scalable uncertainty estimation enables improved molecular property prediction, accelerated optimization, and guided virtual screening.

A novel deep model for efficiently processing 3D pointclouds and fusing with uncertainty for added robustness during sensor failures.

Dual objective learning combined with sparse binary optimization enables jointly positioning discrete sensors and task learning.

A novel continuous-time neuron model with a dynamically changing (liquid) time-constant enables improved representation learning.

Fast, scalable uncertainty estimation algorithm for deep neural networks trained on regression tasks.

A novel bio-inspired neural model enables 100x smaller and more auditable networks for end-to-end autonomous driving.

A Bayesian deep learning approach to soft tactile sensing using GelSight

A class of Neural-ODE's with dynamic time-constants enable emergent "lottery ticket" sparse networks.

Data-driven simulation for RL policy learning. First demonstration of end-to-end vehicles, trained using RL in simulation, directly transferred into the real world.

Dynamic shadow detection allows for agents to detect moving objects earlier than other perception sensors.

A framework for auditing deep networks trained on time-series data.

End-to-end learning algorithm for navigation control and probabilistic pose localization based on visual inputs and noisy GPS.

Learning pose and orientation from visual inputs while capturing uncertainty and rotational ambiguity.

A general algorithm for adaptively re-weighting samples to overcome hidden and harmful dataset biases. Drastically improve existing face detectors.

Variational approach to end-to-end driving along with algorithms for overcoming data imbalances and identifying dangerous OOD edge-cases.

Shadow detection to sense movement even when it is out of sight.

Probabilistic end-to-end (perception-to-control) learning of autonomous vehicles enables greater robustness during deployment.

Learning discrete probabilistic outputs using end-to-end policy networks to estimate continuous control bounds for shared human-robot autonomy.

A Bayesian dropout sampling approach over kernels is amenable to end-to-end control learning from pixel inputs.

Large-scale mathematical modeling of agent behaviors using only existing passive sensors (Bluetooth).

Accelerating the convolution operation enables drastic improvements on timed collision estimation. Results are integrated into JuliaLang.

Mathematical modeling of human mobility using large-scale transaction (purchase) datasets for improve advertisement and urban management.

Using large-scale datasets from passively collected mobile phone records, we discover directional imbalances in human mobility.

Develop mathematical models for human mobility using mobile phone records in industrialized (Portugal) and developing (Ivory Coast) nations. Emergent communities in the underlying network structure provide clues to improving urban design and efficiency.

Tale of two nations: large-scale mobile phone records are used to estimate community network structure differences across nations.

From expert demonstrations, machine learning models are built to continuously estimation motions with 97% accuracy and provide corrective feedback to the user. First place grand-prize at EUCYS and BTYSTE 2011.

From expert demonstrations, machine learning models are built to continuously estimation motions with 97% accuracy and provide corrective feedback to the user. First place grand-prize at EUCYS and BTYSTE 2011.