News

IEEE IRI 2026 was a great experience, hosted this past weekend at UW Bothell. AI was a central theme throughout, most visibly in the keynotes from Bhavani Thuraisingham, Roger Barga, and UW’s John Y. Choe. The program ran two parallel tracks across sixteen sessions plus two workshops, but a few threads showed up consistently. 

AI as a change in how we work, not a replacement for us. This came up often during the keynotes, with what felt like real consensus behind it. The interesting question has shifted from whether AI displaces people to what the reshaped workflow looks like. LLMs, LMMs, and agents dominated much of the research. Many researchers followed a similar pattern: zero-shot evaluations, pre-trained models as feature extractors, large models pointed at a specific problem to see if they’d solve it. Agents were also a consistent theme. For example, we saw MCP-based incident detection, multi-agent recommendation, and RAG got a full session. It was a good overview of what is possible, with interesting use cases and some well documented failure cases. 

Explainability and efficiency also got a lot of attention, but through general-purpose methods. Efficiency leaned on split and federated learning and modality compression, while edge deployment appeared in a few places. The interest is clear, but the specialization to deployment context isn’t yet well characterized. 

Healthcare was another big theme, and the multi-modal work was more heavily emphasized in these scenarios as opposed to explainability. Proteomics with imaging for Alzheimer’s, cross-modal fusion for fundoscopy, modality gating for glioma segmentation, all combining signal types rather than optimizing one, generally with late fusion techniques. 

Overall, the conference was an amazing opportunity and experience to connect with other researchers and share our LI-VAD paper. We fielded great [...]

Intestinal metaplasia (IM) is a histologic finding in gastric and gastroesophageal biopsies, characterized by intestinal-type epithelium and goblet cells. Because IM can appear as small or focal regions, it may be difficult to identify during routine whole-slide review.

We are building an AI-assisted pipeline for detecting IM in H&E-stained whole-slide images. The workflow combines WSI preprocessing, tissue patch extraction, pathologist annotation, model training, validation, and whole-slide inference. The system highlights suspicious IM regions and generates annotation-style outputs, supporting pathologists as a second-reader navigation tool for efficient and sensitive review.

The 2026 IEEE International Conference on Consumer Electronics – Taiwan (ICCE-TW 2026) was held from July 1-3 at South Garden Hotels and Resorts in Taoyuan, Taiwan. ICCE-TW aimed to initiate discussions on research and discovery in electronics and the relevant professional fields. The conference theme was “sustainable and human-centered AI in consumer electronics.” MCL Director, Professor C.-C. Jay Kuo, was invited to give the opening keynote at the conference. His keynote title was “Green Learning for Edge AI: Methodologies and Examples,” with the following abstract.

Rapid AI advances over the last decade have been driven by large amounts of training data and deep learning (DL) technologies. Yet, rising electricity consumption and carbon emissions of AI solutions have been a major concern. Furthermore, the DL decision mechanism is obscure. As an alternative, I have been researching green AI for more than 10 years. The proposed solution, called green learning (GL), does not have neurons, neural networks, or backpropagation. Instead, it uses statistical tools to determine an AI model’s parameters in a feedforward manner. It features smaller model sizes, lower training and testing complexity, transparent mathematics, and a low carbon footprint. Its performance scales well with less training data. GL contains three modules in cascade: 1) representation learning, 2) feature learning, and 3) decision learning. The intuition behind each module and the computational algorithms was discussed. GL offers energy-efficient solutions in data centers and mobile/edge devices. GL has been successfully applied to a few edge AI applications. Several examples were shared to demonstrate its effectiveness and efficiency.

Professor Kuo’s talk was highly praised as inspiring and for its clear vision for future edge AI development.

We are very happy to welcome a new MCL member, Yijin Chen. Here is a quick interview with Yijin:

1. Could you briefly introduce yourself and your research interests?

My name is Yijin Chen. I joined MCL this summer. My main research interest is image recognition based on Green Learning, building mathematically interpretable visual recognition algorithms that can surpass the performance of deep networks without relying on backpropagation. Before joining MCL, I mainly focused on applications of artificial intelligence, as well as some work related to hardware, such as embedded systems and FPGA.

2. What is your impression of MCL and USC?

My first impression of MCL is that it is a very principled lab. Its emphasis on mathematical clarity, reproducible hard metrics, and the requirement that every module can be explained from beginning to end is very different from the “tune until it works” atmosphere I had encountered before. The USC campus itself is also beautiful, and the school’s culture is very appealing. I feel extremely proud to become a member of MCL and of USC.

3. What is your future expectation and plan in MCL?

MCL is a lab that places great emphasis on mathematical thinking, and I hope I can strengthen my own mathematical thinking here. Deep learning and Green Learning are two very different things, and I hope I can approach research on Green Learning from a more rational, more mathematically grounded perspective. We are currently working on advancing Green Learning into its next stage. In the long run, I hope I can make contributions to this work, and grow into a researcher who can explain every design choice from first principles.

Congratulations to Xinyu Wang for passing her defense! Xinyu’s thesis is titled “ Towards Efficient Visual Perception: From Feature Learning to Visual Reasoning.” Here is a brief summary of her thesis:

Visual perception serves as a fundamental component of modern computer vision, enabling the interpretation of large-scale image and video data. As visual data continues to grow in complexity, there is an increasing demand for efficient and scalable frameworks that bridge low-level representations and high-level understanding. This dissertation addresses this challenge by exploring a unified trajectory of efficient visual perception, evolving from feature learning to spatio-temporal modeling and ultimately to scene-conditioned visual reasoning. This dissertation first introduces a statistics-based feature generation framework for image classification, built upon the Least-squares Normal Transform, which reformulates classification as a regression problem for efficient feature learning. It generates discriminative and complementary features, boosting decision learning and training convergence with low computational overhead. It then investigates a particularly challenging visual task, video camouflaged object detection. The proposed GreenVCOD is a lightweight framework that captures temporal context through a Temporal Neighborhood prediction cube, enabling implicit motion modeling without additional computational cost. Building upon this, IDM-VCOD introduces a dual-motion design that combines implicit semantic refinement with explicit motion alignment, along with a selective activation mechanism to balance accuracy and efficiency. Finally, this dissertation shifts toward a reasoning-based paradigm for visual perception. Camouflaged object detection is reformulated as a scene-conditioned pattern-deviation reasoning problem. By leveraging background-aware retrieval and prototype-based reasoning, it identifies subtle deviations without relying on pixel-level supervision. Overall, the proposed methods demonstrate that lightweight design, combined with structured statistical modeling and scene-conditioned reasoning, can effectively address challenging visual perception tasks without relying on heavy supervision or large-scale model training.