News

Professor C.-C. Jay Kuo, Director of MCL, was invited to give a keynote at the AIxMMconference held in Laguna Hills, California, USA, on February 4 (Tuesday). The title ofProfessor Kuo’s keynote is “Mobile/Edge Visual Analytics via Green AI.” The abstract ofis given below.“Mobile/edge visual analytics will prevail in the modern AI era. Most researchers focuson deep-learning-based model compression to achieve this goal. Model compressioncan reduce the model size by 50-80% with slight performance degradation. Modelcompression relies on an existing larger model. The training cost of such a large modelremains. The compression step also demands resources. I have worked on green AIsince 2014, published many papers on this topic, and coined this emerging field “greenlearning.” Green learning demands low power consumption in both training andinference. It has attractive characteristics, such as small model sizes, fewer trainingsamples, mathematical transparency, ease of incremental learning, etc. It can reducethe model size of its deep-learning counterpart by 95-99%. The training can beconducted from scratch. The resulting model is inherently smaller. It is ideal for mobileand edge devices. Green learning relies on signal-processing disciplines such as filterbanks, linear algebra, subspace learning, probability theory, etc. Although it exploitsoptimization, it avoids end-to-end system optimization, a non-convex optimizationproblem. Instead, it adopts modularized optimization, and each optimization problemcan be cast as convex optimization. In this example, I will use several examples todemonstrate the advantages of green learning in visual analytics for mobile/edgedevices.”Professor Kuo received quite a few questions immediately after the talk. He alsoparticipated a panel discussion in the afternoon, 1-2:30 pm.

Congratulations to Ganning Zhao for passing her defense. Ganning’s thesis title is “Learning to Generate Better: Visual Refinement and Evaluation in Generative AI Models.” Her Dissertation Committee included Jay Kuo (Chair), Antonio Ortega, and Stefanos Nikolaidis (Outside Member). The Committee members were pleased with the quality of Ganning’s thesis work. Thanks to our lab members for participating in her rehearsal and providing valuable feedback.Here is the abstract of Ganning’s thesis:

Generative artificial intelligence (GenAI) has advanced rapidly in recent years, finding widespread applications in numerous domains. Generative models (GMs), which produce new data by learning underlying data distributions, typically require vast quantities of training examples—often in the millions or billions. Acquiring and labeling such large datasets is expensive and labor-intensive, prompting researchers to use synthetic data for training. However, the limited quality of these synthetic samples can restrict model performance. Furthermore, accurately evaluating the quality of generated images is critical for guiding improvements in generative models. This dissertation addresses two core challenges: (1) refining synthetic images, and (2) evaluating the quality of generated samples. Each challenge is tackled with a foundational approach and its subsequent enhancement.

Welcome to 2025, MCL family! Here’s to a year full of happiness, success, and the courage to chase your dreams. May this be a time of growth, discovery, and opportunities that take you to new heights. Let’s make it a year to remember—graceful, bold, and uniquely ours. Cheers to the journey ahead!

Photo credits: https://www.freepik.com/

Pathology imaging is a key method in medical diagnostics, offering detailed information on tissue structures and biomarkers. However, traditional approaches to measuring biomarker intensity can be time-consuming, which is why new machine-learning methods are needed to predict these biomarkers directly from images. By identifying and analyzing biomarkers in this way, doctors can better predict patient grades, postoperative responses, and overall disease progression. This approach also presents significant commercial potential.

To address this issue, we adopted the Green U-Shaped Learning (GUSL) pipeline. Specifically, we extracted 15 randomly selected regions of interest (ROIs) from each pathology slide and resized them into patches. Each ROI was then analyzed to determine its biomarker density. GUSL utilizes Green Learning techniques like PixelHop, RFT, and LNT to integrate features from different spatial scales and refine predictions level by level. While this method shows promising results, it requires further adjustments to effectively handle ROIs with larger spatial sizes.