Summer 2022 started on Wednesday, May 18, last week. As Covid-19 is more well-controlled nowadays, we gradually resumed the lab weekly activities on campus. Hope everyone have a great summer time!
Image credit:
https://alumni.usc.edu/campus/
https://www.morainevalley.edu/summer22/
https://www.linkedin.com/school/university-of-southern-california-viterbi-school-of-engineering/
Five MCL members attended the Viterbi PhD hooding ceremony on Wednesday, May 11, 2022, 8:30-11:00 a.m. in the Bovard Auditorium. They were Mozhdeh Rouhsedaghat, Tian Xie, Yijing Yang, Kaitai Zhang, and Min Zhang. Congratulations to them for their accomplishments in completing their PhD program at USC!
Mozhdeh Rouhsedaghat received her Bachelor’s in Electrical Engineering from the Sharif University of Technology in 2017 and then joined the University of Southern Californian for her Ph.D. studies in Fall 2017. She received her Master’s and Ph.D. in Electrical Engineering from the University of Southern California in 2021 and 2022, respectively. Her research interests include computer vision, vision-and-language, and adversarial learning. Her Ph.D. thesis title is “Data-Efficient Image and Vision-and-Language Synthesis and Classification”.
Tian Xie received his B.S. degree in Physics from Fudan University, Shanghai, China, in 2017. He then joined the University of Southern California (USC) as a Ph.D. student. His research interest is graph learning, machine learning, and data mining. His thesis title is “Efficient Graph Learning: Theory and Performance Evaluation”. He will join Meta as a Research Scientist.
Yijing Yang received her Bachelor’s degree in 2016 from Tianjin University, China, and received her Master’s degree in Electrical Engineering from USC in 2018. She then joined MCL as a PhD student guided by Prof. C.-C. Jay Kuo. Her research interests include image processing, computer vision, and medical image analysis.
Kaitai Zhang defended his Ph.D. in Electrical and Computer Engineering and graduated from University of Southern California in May 2021, where he is fortunate enough to be advised by Professor C.-C. Jay Kuo. During his PhD study, Kaitai conducted research projects from image processing to computer vision using various machine learning and deep learning techniques. Before that, he obtained his bachelor’s [...]
Congratulations to Tian Xie for passing his defense on May 4, 2022! His Ph.D. thesis is entitled “Efficient Graph Learning: Theory and Performance Evaluation”. Here we invite Tian to share a brief introduction of his thesis and some words he would like to share at the end of the Ph.D. study journey.
1) Abstract of Thesis
Graphs are generic data representation forms that effectively describe the geometric structures of data domains in various applications. Graph learning, which learns knowledge from this graph-structured data, is an important machine learning application on graphs. In this dissertation, we focus on developing efficient solutions to graph learning problems. In particular, we first present an advanced graph neural network (GNN) method specified for bipartite graphs that is scalable and without label supervision. Then, we investigate and propose new graph learning techniques from the aspects of graph signal processing and regularization frameworks, which identify a new path in solving graph learning problems with efficient and effective co-design.
From the GNN perspective, we extend the general GNN to the node representation learning problem in bipartite graphs. We propose a layerwise-trained bipartite graph neural network (L-BGNN) to address the challenges in bipartite graphs. Specifically, L-BGNN adopts a unique message passing with adversarial training between the embedding space. In addition, a layerwise training mechanism is proposed for efficiency on large-scale graphs.
From the graph signal perspective, we propose a novel two-stage training algorithm named GraphHop for the semi-supervised node classification task. Specifically, two distinctive low-pass filters are respectively designed for attribute and label signals and combined with regression classifiers. The two-stage training framework enables GraphHop scalable to large-scale graphs, and the effective low-pass filtering produces superior performance in extremely small label rates.
From the regularization framework perspective, we [...]
Facial Expression Recognition (FER) is a challenging topic in the image classification field. Some of its applications, such as driver assistance systems, require real-time response or demand methods that can run on low-resources devices. FER can be classified into conventional methods and deep learning methods. Deep learning-based methods have attracted much attention in recent years because of their higher performance even under challenging scenarios. However, deep learning-based methods rely on models that demand high computational resources. At the same time, conventional methods depend on hand-crafted features that may not perform well in different scenarios. In this context, some studies are pursuing to reduce the computational complexity of deep learning models while achieving similar results to those more complex models. But even these models with reduced complexity can require a lot of computational resources.
To tackle this problem, we propose ExpressionHop. ExpressionHop is based on a Successive Subspace Learning classification technique called PixelHop[1], which allows us to automatically extract meaningful features without the need for higher computational demanding models. As shown in Figure1, we first extract facial landmark patches from face images, and then use Pixelhop to extract feature. Discriminant feature test is utilized for feature selection before doing classification using logistic regression. As shown in table1, our model achieved higher or similar results compared to traditional and deep learning methods for JAFFE, CK+, and KDEF datasets. At the same time, a comparison of the number of parameters of the models indicates that the proposed model demands fewer computational resources even when compared to newer deep learning methods that rely on reduced complexity models.
— By Chengwei Wei and Rafael Luiz Testa
MCL Director, Professor C.-C. Jay Kuo, was selected to receive a Phi Kappa Phi Faculty Recognition Award by the honor society of the Phi Kappa Phi chapter at the University of Southern California. His achievement was recognized at the annual USC Academic Honors Convocation at Town and Gown on April 19,2022.
Phi Kappa Phi presented up to four awards to USC faculty for works predominantly created while at USC. The recipients of the awards are chosen by a committee of faculty from diverse fields. The prize recognizes major achievements in research, contributions to knowledge, and artistic creativity. Scholarly and creative works. Professor Kuo was selected because of his collection of work on interpretable A.I. and sustainable computing in recent years.
The following excerpt is taken from the description that commemorates Professor Kuo’s Phi Kappa Phi Faculty Recognition Award. “Professor Kuo’s articles on interpretable and green artificial intelligence have received far-reaching recognition. With this work, he creates new algorithms for A.I. that use far less computational processing power and reduce energy consumption in ways that make this technology more sustainable. His artificial intelligence and machine learning research also has many promising applications for the internet, media, and national security. This work has earned him accolades, as his research on interpretable convolutional neural networks received a Best Paper Award from the Journal of Visual Communication and Image Representation.”
A progressive edge-guided generative model, called ProGenHop, is presented in this work.
The majority of existing generative models utilize neural networks to represent the underlying data distribution. Alternatively, ProGenHop offers a novel method based on Successive Subspace Learning (SSL) for feature extraction and image generation. The benefits of ProGenHop are its interpretability and significantly lower computational complexity, as opposed to compulationally-complex, black-box neural networks. ProGenHop maintains generation quality with a small training size. Moreover, ProGenHop is easily extendable to further generative model applications, such as attribute-guided image generation, super resolution, and high-resolution image generation.
A generative model learns a distribution for the underlying dataset during the training phase. During the generation phase, samples can be drawn from the distribution as new data. Most of the prior work like the GAN-based, VAE-based, and diffusion-based generative models utilize neural networks to learn complex non-linear transformations. In our work, we utilize the SSL pipeline for feature extraction. An SSL pipeline consists of consecutive Saab transformations. In essence, the SSL pipeline receives an RGB image and converts it into a feature vector. Since Saab transformation is a variant of Principle Component Analysis (PCA), it inherits PCA properties; One of these nice properties is that the Saab transformation generates feature vectors with uncorrelated components. This property facilitates the utilization of Gaussian priors for generative model training.
ProGenHop is an unconditional generative model which has a progressive approach in generating images: it starts the unconditional generation in a low-resolution regime, then sequentially increases the resolution via a cascade of conditional generation modules. ProGenHop has three modules, namely, Generation Core, Resolution Enhancer, and Quality Booster. The first module learns the distribution of low-resolution images using a Gaussian mixture model and performs unconditional image [...]
Chronic diseases like chronic obstructive pulmonary disease (COPD) and asthma have high prevalence and reduce the compliance of the lung, thereby impeding normal ventilation. Functional lung imaging is of vital importance for the diagnosis and evaluation of these lung deseases. In recent years, high performance low field systems have shown great advantages for lung MRI imaging due to reduced susceptibility effects and improved vessel conspicuity. These MRI configurations provide improved field homogeneity compared with conventional field strengths (1.5T, 3.0T). More possibilities are brought to the researchers to detect regional volume changes throughout the respiratory cycle at lower field strengths, such as 0.55T.
Recently, under the collabration between Dynamic Imaging Science Center (DISC) and MCL, an approach for regional functional lung ventilation mapping using real-time MRI has been developed. It leverages the improved lung imaging and improved real-time imaging capability at 0.55T, without requiring contrast agents, repetition, or breath holds. In the image acquisition, a sequence of MRI in the time series representing several consecutive respiratory cycles is captured. To resolve the regional lung ventilation, an unsupervised non-rigid image registration is applied to register the lungs from different respiratory states to the end-of-exhalation. Deformation field is extracted to study the regional ventilation. Specifically, a data-driven binarization algorithm for segmentation is firstly applied to the lung parenchyma area and vessels, separately. A frame-by-frame salient point extraction and matching are performed between the two adjacent frames to form pairs of landmarks. Finally, Jacobian determinant (JD) maps are generated using the calculated deformation fields after a landmark-based B-spline registration.
In the study, the regional lung ventilation is analyzed on three breathing patterns. Besides, posture-related ventilation differences are also demonstrated in the study. It reveals that real-time image acquisition [...]
With the pervasive of image steganography algorithms in social media, image steganalysis becomes inevitably important nowadays. One of the most secure steganographic scheme is called content-adaptive steganography. WOW, S-UNIWARD, HUGO and HILL are all successful steganography algorithms of this kind. Since content-adaptive steganography will calculate embedding cost after they evaluate the cover image, and will tend to put more embeddings in complex regions. It makes it harder for image steganalyzers to detect if the image has been embedded information or not.
Our goal is to provide a data-driven method, which does not apply hand-crafted high-pass filtering in preprocess step or any neural network based architectures. We have unsupervised feature extraction and machine learning-based classifier to fulfill the task. Specifically, we first split input image into 3×3 blocks, and partition blocks into several groups based their embedding cost. We use Saab transform to extract features on blocks and make decision. The difference of soft decision scores on cover image and stego image are efficient for us to do image-wise decision. In order to find the embed locations in unseen images, we train an embed location classifier from block soft decision scores, as shown in Fig. 1. Based on embed location probability score from each group, we train the final image-wise ensemble classifier and give us the image-level decision, as shown in Fig.2 .
Compared to CNN-based steganalysis models, our method does not use end-to-end training and backward propagation. Therefore, it is very light-weight in terms of model size and memory usage. In the meantime, our method can beat all traditional steganalysis method and some benchmarking CNN-based model.
— by Yao Zhu
A robust fake satellite image detection method, called Geo-DefakeHop, is proposed in this work. Geo-DefakeHop is developed based on the parallel subspace learning (PSL) methodology. PSL maps the input image space into several feature subspaces using multiple filter banks. By exploring response differences of different channels between real and fake images for filter banks, Geo-DefakeHop learns the most discriminant channels and uses their soft decision scores as features. Then, Geo-DefakeHop selects a few discriminant features by validation dataset from each filter bank and ensembles them to make a final binary decision. Geo-DefakeHop offers a light-weight high-performance solution to fake satellite images detection. Its model size is analyzed, which ranges from 0.8 to 62K parameters. Furthermore, it is shown by experimental results that it achieves an F1-score higher than 95% under various common image manipulations such as resizing, compression, and noise corruption.
— By Max Chen
Graph-based semi-supervised learning has shown prominent performance in node classification task by exploiting the underlying manifold structure of data. Recently, an enhancement on the classical label propagation (LP) named GraphHop is proposed, which has outperformed the existing graph convolutional networks (GCNs) on various networks. Although the superior performance in GraphHop model is explained in the view of smoothening both node attribute and label signals, its mechanisms are still not fundamentally clear.
In this work, we develop deeper insights into the the GraphHop model from the point of regularization framework. We show that GraphHop model can be cast into an iterative approximated optimization of a particular regularization function on graphs. Then, based on this variational interpretation, we propose two approaches to address the limits in the GraphHop model due to the approximated optimization process. In particular, these are 1) additional aggregations in optimizing the label embeddings; 2) adaptively selecting of the reliable unlabeled samples for the classifier training. Experiments show that equipped with these two improvements, our model called GraphHop++ is able to gain significantly better performance than the former GraphHop model, in addition to the state-of-the-art methods on various benchmark networks with limited label rates.
— By Tian Xie
