Radoslav Ivanov
Rensselaer Polytechnic Institute, NY·Computer Science
Self-driving cars, farm robots and medical devices increasingly use neural networks (NNs) to interpret raw sensor data such as camera images. The NN output then feeds traditional 'symbolic' components, such as state estimators, which track where the system is, and controllers, which decide how it should act.
- Fields
Computer Science
Engineering
- Topics
cyber-physical systems
neuro-symbolic methods
neural network calibration
- Grants
- 1 known grant, through May 2027
What they work on
- Develop extrinsic neuro-symbolic calibration that adjusts the uncertainty of already-trained neural networks to match the assumptions of downstream state estimators and controllers, without retraining the networks.
- Develop intrinsic neuro-symbolic training that trains and calibrates neural networks together by combining standard calibration losses with symbolic losses tied to the downstream estimation or control task.
- Make calibration robust to temporary sensor faults and consistent with system dynamics over time, for both probabilistic and bounded-input assumptions, and demonstrate it on the 1/10-scale F1/10 autonomous racing car.
More about the research
These downstream components rely on assumptions about how uncertain their inputs are. NNs can break those assumptions, because even small input changes or shifts in data can distort their outputs. That mismatch can undermine the safety guarantees of the whole system.
Why it matters
Neural network perception is spreading into transportation, agriculture and medicine, where errors can be dangerous. Calibration methods that keep NN outputs consistent with the assumptions of estimators and controllers could make these systems safer and more predictable. They could also let engineers use proven tools such as resilient state estimation and robust control in camera-driven machines.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Sayan Mitra
University of Illinois Urbana-Champaign, IL·Computer Science
Develop refinement algorithms that synthesize perception contracts, machine-checkable bounds on how wrong an AI perception module can be, together with the operating design domains in which they hold, for pipelines that fuse camera and inertial sensor data.
Nikolai Matni
University of Pennsylvania, PA·Engineering
Develop safety-rich data augmentation and domain randomization techniques for training learning-enabled systems, and determine which types of data must be collected to guarantee end-to-end safety.
Qi Chen
University of California-Irvine, CA·Computer Science·Early-career
Develop system-to-AI and AI-to-system mapping methods that generate physical-layer attacks on autonomous vehicle AI while keeping them physically realizable and semantically equivalent across diverse real system designs.
The school
Rensselaer Polytechnic Institute: 5 labs on PI Pages, $140M a year on research, 113 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.