News

The 2025 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR) was held in San Jose from August 6 to 8. The event commenced with three keynote addresses: Prof. Prem Devanbu (University of California, Davis) discussed the reliability of large language models for code generation and offered guidance on when their results can be trusted. Prof. Edward Y. Chang (Stanford University) presented adaptive multi-modal learning as a way to address LLM limitations. Dr. Ed H. Chi (Google DeepMind) spoke on the future of AI-assisted discovery, highlighting systems that enhance rather than replace human expertise.During the conference, Mahtab Movhhedrad, a member of the Media Communications Lab (MCL), presented the paper “GUSL-Dehaze: A Green U-Shaped Learning Approach to Image Dehazing.” This work introduced GUSL-Dehaze, a physics-based green learning framework for image dehazing that completely avoids deep neural networks. The method begins with a modified Dark Channel Prior for initial dehazing, followed by a U-shaped architecture enabling unsupervised representation learning. Feature-engineering techniques such as Relevant Feature Test (RFT) and Least-Squares Normal Transform (LNT) were employed to keep the model compact and interpretable. The final dehazed image is produced through a transparent supervised learning stage, allowing the method to achieve performance comparable to deep learning approaches while maintaining a low parameter count and mathematical transparency.The conference also included a panel session, “Learning Beyond Deep Learning (LB-DL) for Multimedia Processing,” chaired by Prof. Ling Guan (Toronto Metropolitan University/Ryerson University) and Prof. C.-C. Jay Kuo (University of Southern California, Director of the MCL Lab). Prof. Kuo discussed emerging paradigms that challenge the dominance of deep learning, emphasizing the growing importance of interpretability, efficiency, and sustainability in shaping the next generation of multimedia research.

Magnetic Resonance Imaging(MRI) is a non-invasive, radiation-free scanning procedure generally used to obtain images of internal organs. This imaging modality is a popular screening technique for Prostate Cancer. A Prostate MRI may be used to detect prostate cancer, determine the requirement for biopsy, guide needles during targeted MRI biopsy, or detect the spreading of cancer to neighboring areas. Current standards of MRI acquisition still lead to errors, like false detections, where patients are unnecessarily sent to biopsy, or missed detections, where existing tumors remain undetected. A good quality MRI is essential to ensure that Prostate Cancer is diagnosed on time.

An MRI has several sequences, namely T2W, ADC, and DCE. For an MRI to be of diagnostic quality, atleast two of the three sequences must independently be of diagnostic quality. Errors/artifacts may be present in some or all of these sequences. Some common artifacts include motion artifacts, rectal gas, hip prosthesis, etc., that affect the quality of the MRI.

In order to assess the quality of an MRI, we train independent models for each of these sequences. While MRI images are volumetric, we treat each slice of the MRI independently. 2D Haar Wavelet Transform is applied to extract features from the LL, LH, HL, and HH bands. These features are extracted at two different resolutions. The Discriminant Feature Test(DFT) is used to reduce the feature dimension by removing features with a high DFT loss. An XGBoost Classifier is then trained using these selected features to predict whether each slice of the MRI is of satisfactory quality or not. The quality predictions from each of the three sequences are then combined to obtain the final MRI quality prediction. This approach is light-weight, efficient, and explainable, with [...]

Wavelet-based Green Learning (GreenWave) is a new image classification framework that combines the multi-scale power of wavelet transforms with the efficiency and interpretability of Green Learning principles. It avoids backpropagation and replaces traditional deep learning architectures with a transparent, feedforward pipeline.

At its core, GreenWave begins by applying a discrete wavelet transform (DWT)—typically using Haar wavelets—to each image, capturing both local and global spatial structures at multiple resolutions. It extracts features from wavelet subbands (like LL, LH, HL, HH) as well as local image patches from different regions (e.g., east, south, west, north). These features are used to construct class templates via averaging over training examples.

GreenWave operates in three rounds of classification:1. Round-1 classifies easy samples using cosine similarity between the input and class templates (one-vs-rest). It uses confidence metrics (like entropy) to decide which samples are “easy” and can exit the pipeline early.2. Round-2 focuses on semi-hard samples using updated templates and discriminant feature test (DFT) masks to emphasize class-informative coefficients. It also introduces one-vs-one templates to resolve class confusion.3. Round-3 targets the hardest samples using further refined templates and updated confusion masks, maximizing accuracy while maintaining interpretability.Throughout all stages, GreenWave uses cosine similarity as its feature-matching metric and XGBoost classifiers for decision learning, completely bypassing gradient-based training.

Overall, GreenWave demonstrates that a non-backpropagation, wavelet-template-based system can achieve near state-of-the-art performance while being highly efficient, explainable, and modular. This makes it an ideal choice for low-resource or transparent AI applications.

XGBoost is a widely used gradient-boosting framework in machine learning. Due to its low resource consumption and interpretability for sequential learning, XGBoost has gained increasing attention in green learning, where computational efficiency and model transparency are essential. However, XGBoost is not optimized in the current green learning literature. It accounts for a large part of the model parameters in a green learning system. It is desired to reduce the size of XGBoost while maintaining its high performance.

We recently have proposed a new method called Multi-stage XGBoost (MXB), a modular framework that builds a sequence of XGBoost models. MXB adopts a sequence of shallow XGBoost models, each of which defines a stage. Unlike conventional XGBoost, which is trained on the full feature set, the model at each stage of MXB is trained on a distinct feature subset. The general idea is to put more discriminant ones in earlier stages and less discriminant ones in later stages. This is motivated by the convergence curve of the XGBoost classifier and regressor. Also, the transition from one XGBoost model to the next in MXB is governed by the behavior of the gradient and Hessian, which represent the first- and second-order derivatives of the objective function, respectively. The gradient quantifies the direction and rate of change of the loss function, indicating how far the model is from optimality. The Hessian, on the other hand, measures the curvature of the loss, offering insight into the local stability of convergence.

We evaluate MXB on two datasets: MNIST and Fashion-MNIST, and compare it with standard XGBoost models of equivalent depth and model size. The training and testing curves show that MXB has smaller train–validation gaps and more stable testing loss as compared to [...]