News

Supervised feature learning is a critical component in machine learning, particularly within the green learning (GL) paradigm, which seeks to create lightweight, efficient models for edge intelligence.

Supervised feature learning involves creating new, more discriminant features from an existing set of features, typically produced during an earlier stage of unsupervised representation learning. The objective is to enhance the discriminative power of the features, thereby improving the model’s accuracy and robustness in decision-making tasks such as classification.

The process typically begins with a rich set of representations obtained from the unsupervised module. These representations are then rank-ordered according to their discriminant power using a discriminant feature test (DFT). However, if the initial set of features lacks sufficient discriminant power, new features can be derived through linear combinations of the existing ones. This method transforms a multi-class classification problem into multiple binary classification problems, then applies linear regression to generate new features that are more discriminant than the original ones. These new features are shown to improve the performance of the classifier, demonstrating the effectiveness of the supervised feature learning process within the GL framework.We proposed the Least-Squares Normal Transform (LNT) for generating new discriminant features. This method transforms a multi-class classification problem into multiple binary classification problems, then applies linear regression to generate new features that are more discriminant than the original ones. These new features are shown to improve the performance of the classifier, demonstrating the effectiveness of the supervised feature learning process within the GL framework.

References:

X. Wang, V. K. Mishra, and C.-C. J. Kuo, “Enhancing edge intelligence with highly discriminant lnt features,” in 2023 IEEE International Conference on Big Data (BigData). IEEE, 2023, pp. 3880–3887

Digital cameras typically use a color filter array (CFA) over the image sensor to capture color images, with the Bayer array being the most common CFA. This array captures only one color per pixel, resulting in raw data that lacks two-thirds of the necessary color information. Demosaicking is the process used to reconstruct the complete color image from this partial data. Simple interpolation methods like bilinear and bicubic often produce suboptimal results, especially in complex areas with textures and edges.

To improve image quality, adaptive directional interpolation methods align the interpolation with image edges, using techniques like gradients or Laplacian operators to detect horizontal and vertical edges. Recently, deep learning techniques have set new performance benchmarks in demosaicking, but their complexity and resource demands pose challenges for deployment on edge devices with limited processing power and storage.

To address these issues, a lightweight “green learning” approach is proposed for demosaicking on edge devices. Unlike traditional deep learning models, green learning does not rely on neural networks. Our proposed model can be explained in three stages. See Fig [1] for details. beginning with data processing, where the interpolation method is used to estimate RGB values. Channels are then categorized into subchannels based on their positions in the CFA array, improving prediction accuracy in the learning stage. fig[2] for details. In the feature processing stage of the green learning approach for demosaicking, three subsequential modules work together to enhance the model’s performance. First, unsupervised representation learning techniques, such as the Saab transform or Successive Subspace Learning (SSL), are used to efficiently extract meaningful representations from raw data. Next, supervised feature selection is performed using the Discriminant Feature Test (DFT) and the Relevant Feature Test (RFT) to identify and [...]

Research in healthcare systems has been mainly focused on automating several tasks in the clinical pipeline, aiming at enhancing or expediting physician’s diagnosis. Prostate Cancer (PCa) is widely known as one of the most frequently occurring cancer in diagnosis in men. If early diagnosed, the mortality rate is almost zero. Yet, should it goes under the radar -in the metastasis stage- that rate plummets at 31% [1]. For diagnosis, after a high level of Prostatic Specific Antigen (PSA), patients are recommended to undergo an MRI screening. Those patients with suspiciously looking lesions on the prostate gland will eventually undergo biopsy. The histology gives the definite answer whether a patient suffers from cancer. Nevertheless, it is observed in real practice that urologists tend to over-diagnose patients with csPCa, thus increasing the number of unnecessary biopsies. As such, this increases the diagnostic costs and patient discomfort.

Computer vision empowered with AI has shown promising results in the last decade. Computer-Aided Diagnosis (CAD) tools benefit from AI’s rapid evolution and many works have been proposed to automatically perform lesion detection and segmentation. Even though the Deep Learning (DL) paradigm is ubiquitous in modern AI, medical applications require more transparency behind feature extraction and thereby DL is often deemed as a “black-box” from physicians. Our proposed pipeline, PCa-RadHop [2], employs a novel and linear module for data-driven feature extraction and hence the decision making becomes more interpretable. PCa-RadHop receives three different modalities as input from the MRI-scanner (i.e. T2w, ADC, DWI), pertinent to PCa diagnosis. It consists of two stages. The first stage calculates a probability map about csPCa presence in a voxel-wise manner, while the second stage is meant to reduce the false positive rate on that heatmap [...]

Professor C.-C. Jay Kuo, Director of MCL, attended the IEEE Conference on Multimedia Exposition (ICME) held in Niagara Falls, Canada, from July 15-19, 2024. Professor Kuo had dual roles in this conference as a Panel co-Chair and a keynote speaker. Professor Kuo gave his keynote on 7/18 (Thursday) on “Toward Interpretable and Sustainable AI via Green Learning.” Besides, Dr. Kuo and Dr. Zicheng Liu of AMD organized a panel as summarized below.

Panel Title: Generative AI – Opportunities, Challenges, and Open Questions

Panel Background: Generative AI has received a lot of attention due to the tremendous success of ChapGPT. Large foundation models have been trained, leading to various demos and potential applications such as text-to-image and text-to-video cross-domain generations. Resources have been invested in building massive computational and storage infrastructures. Furthermore, data collection and cleaning are essential to high system performance. In the face of these rapid developments, this panel will discuss opportunities, challenges, and open questions associated with generative AI.

Four panelists were invited:

Rogerio Feris, IBM Research

Lijuan Wang, Microsoft Research

Jiebo Luo, University of Rochester

Junsong Yuan, State University of New York at Buffalo

Q&A topics:

Today’s generative AI is tilted more toward “engineering” than “science.” Will this be a concern in the long run?

What are the major shortcomings of the current large foundation models?

How vital are “data collection and cleaning” tasks in generative AI? How do large companies carry out such tasks? Will we run out of data? If so, how soon?

Will “copyright,” “plagiarism,” and “hallucination” be issues? How can we address them? How can we trust the answers?

What roles can small AI companies and academia with limited resources play?

What is the future R&D direction of generative AI? What will be the next big breakthroughs?

Dr. Kuo also had a [...]

Magnetic resonance imaging (MRI) is a good way to detect clinically significant prostate cancer and guide biopsies, due to the superior resolution and contrast of imaging, without harming the human body. Based on prostate MRI, prostate segmentation is a process to localize prostate boundaries for radiotherapy and automate the calculation of the prostate volume. Automatic prostate segmentation is an important step in computer-aided diagnosis of prostate cancer and treatment planning [1].

It is very hard to collect and obtain large annotated datasets for AI In Healthcare. We worked with USC Keck Medical School on this project, and they provided us with a large medical dataset, which was very precious and helpful. In addition to this dataset, we also used some public datasets like ISBI-2013 [2] and PROMISE-12[3] to analyze and evaluate our Green U-Shaped Learning (GUSL) methodology.

Our Green U-Shaped Learning (GUSL) framework is a feed-forward encoder-decoder system based on successive subspace learning (SSL), and it consists of two modules: 1) encoder: fine to coarse unsupervised representation learning with cascaded VoxelHop units, and 2) decoder: coarse to fine segmentation prediction with voxel-wise regression and local error correction. Our model is lightweight and totally transparent while keeping comparable performance.

We have done 5 cross-validations for the dataset from USC Keck Medical School. For T2-cube MRIs, the Dice Similarity Coefficient (DSC) of the prostate segmentation was over 93%. The USC Keck Medical School doctors were very satisfied with these impressive results. In the next step, we will apply our GUSL model to some public datasets and then compare and analyze the performance of our method with some state-of-the-art Deep Learning methods. In the future, we aim to develop methods for segmenting other organs like cardiac. I hope our methods [...]