Zhuolin Jiang
Principal Research Scientist · RTX

Research

My research focuses on Computer Vision, Pattern Recognition and Machine Learning, specifically on the following topics:

Topics

Submodular Optimization for Vision

Submodularity is an intuitive diminishing returns property, stating that adding an element to a smaller set helps more than adding it to a larger set. Submodularity allows one to efficiently find (near-)optimal solutions, which is useful in a lot of vision applications. Our research aims to use submodularity optimization to solve various vision problems.

Submodular Saliency
Submodular Dictionary Learning
Action Attributes

Sparse Coding and Dictionary Learning

Sparse coding approximate an input signal as a linear combination of a few items from a predefined and learned dictionary. It usually achieves state-of-the-arts in all kinds of vision applications. The performance of sparse coding relies on the quality of dictionary. Our research aims to learn a discriminative dictionary for recognition.

LC-KSVD
Discriminative Dictionary Learning with Pairwise Constraints
Online Semi-supervised Discriminative Dictionary Learning
Tag Taxonomy Aware Dictionary Learning
Discriminative Tensor Sparse Coding

Low-Rank Matrix Recovery for Vision

A common modeling assumption in many applications is that the underlying data lies (approximately) on a low-dimensional linear subspace. That is, a matrix X can be decomposed into two matrices: X = A+E, where A is a low-rank matrix and E is a sparse matrix. Low-rank matrix recovery which determines the low-rank matrix A from X, has been successfully applied to many applications. Our research aims to use this technique for multi-class classification.

Low-rank Matrix Recovery

Unsupervised and Supervised Clustering

Data clustering is an important task in vision. I used it to learn action prototypes (or action prototype tree). A large number of studies aim to improve clustering by using supervision in the form of pairwise constraint or category information of each point. We used the category information to enforce discriminativeness for each cluster so the final clusters good for classification.

Unified Tree-based Framework
Class Consistent k-means

Transfer Learning

Many learning approaches work well only under a common assumption: training and testing data are drawn from the same feature space and distribution. In many practical applications, the assumption may not hold. In such cases, transfer learning between task domains would be desirable since it is expensive to recollect training data and rebuild the model. Our research aims to transfer knowledge across domains and transfer from multiple such source domains.

Transferable Dictionary Pair
View-invariant Sparse Representations

Efficient LLM Reasoning

Reasoning-capable LLMs often produce excessively long chains with redundant or uninformative steps, increasing inference cost and sometimes harming accuracy and safety. Our research aims to improve efficiency by pruning uncritical reasoning paths to produce shorter and more reliable reasoning trajectories.

frost

Out-of-Distribution Detection

Out-of-distribution (OOD) detection is critical for building reliable AI systems, particularly in high-stakes and open-world settings. Recent work has explored feature-shaping approaches that transform neural representations to separate in-distribution and OOD samples; however, such methods often generalize only to specific datasets or model architectures. Our work aims to derive more generalizable OOD features from an information-theoretic perspective.

OOD features