News

Word embeddings, also known as distributed word representations, learn real-valued vectors that encode words’ meaning. They have been widely used in many Natural Language Processing (NLP) tasks, such as text classification, part-of-speech tagging, parsing, and machine translation. Text classification is a task where the input texts have to be classified into different categories based on their content. Word embedding methods have been tailored to text classification for performance improvement.

In this research, two task-specific dependency-based word embedding methods are proposed for Text classification. In contrast with universal word embedding methods that work for generic tasks, we design task-specific word embedding methods to offer better performance in a specific task. Our methods follow the PPMI matrix factorization framework and derive word contexts from the dependency parse tree. As compared linear contexts, dependency-based contexts can find long-range contexts and exclude less informative contexts. One example is shown in Fig. 2, where the target word is ‘found’. Guided by the dependency par-sing tree, its closely related words (e.g. ‘he’, ‘dog’) can be easily identified. In contrast, less related words (e.g. ‘skinny’, ‘fragile’) are gathered by linear contexts.

Firstly, to construct robust and informative contexts, we use dependency relation which represents the word’s syntactic function to locate the keywords in the sentence and treat the keywords and the neighbor words in the dependency parse tree as contexts.

To further increase the text classification performance, we make our word embedding learns from word-context as well as word-class co-occurrence statistics. We combine the word-context and word-class mutual information into a single matrix for factorization.

It is shown by experimental results they outperform several state-of-the-art word embedding methods.

Image credits:

Image showing a simple example algorithm framework for text classification is from https://laptrinhx.com/nlp-multiclass-text-classification-machine-learning-model-using-count-vector-bow-tf-idf-2622024659/

In Fall 2021, we have a new MCL member, Qingyang Zhou, joining our big family. Here is a short interview with Qingyang with our great welcome.

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

Hello, I am Qingyang Zhou, a new member of MCL in 21fall. I received my Bachelor’s and Master’s degrees both in Shanghai, China. My research interest is multimedia signal processing, especially video coding and point cloud compression. I am passionate about these research areas and I have been working on different video coding standards for over 3 years before coming to USC. I am very happy to share with you my experience in this field.

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

MCL is filled with productive and professional researchers. The lab offers us many different research directions from multimedia signal processing to computer vision. Professor Kuo is creative, hardworking, and full of insightful thoughts. I believe I would receive excellent academic training and become a good researcher at MCL. USC is also an excellent place with both a beautiful campus and strong academic background. I am very happy that I could continue my research at USC.

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

At MCL, I will continue to do research on multimedia-related areas. I have been doing this for years and there are still many questions waiting for me to solve. For the next few months, my attention will be mainly focused on Point Cloud Compression. For the next several years, I would like to get my academic skills improved and make good preparations before entering the industry or academia.

An image generative model learns the distribution of image samples from a certain domain and then generates new images that follow the learned distribution.  The design of image generative models involves analysis and generation two pipelines.  The former analyzes properties of training image samples while the latter generates new images after the training is completed. Only the generation unit is used for image generation in inference.  There is a resurge of interests in generative models due to the amazing performance achieved by deep-learning-based (DL-based) methods in general and generative adversarial networks (GANs) in particular.  Yet, DL-based methods attempt to solve a nonconvex optimization problem which is difficult to explain. GAN’s training may suffer from gradient vanishing, convergence difficulty and mode collapse. Furthermore, its implementation demands higher computational resources due to large model sizes.

The design of GANs demands that distributions of training and generated images be indistinguishable, which is implicitly achieved by training a generator/discriminator pair through end-to-end optimization of a cost function.  In contrast with the GAN approach, we propose a novel and explainable image generation method with explicit sample distribution modeling in this work.  For image analysis, we construct fine-to-coarse spatial-spectral subspaces using the PixelHop++ architecture and obtain sample distributions in each subspace. For image generation, we reverse the process by generating samples in the coarsest subspace and add more details to these samples gradually. Our solution, called GenHop (an acronym of Generative Pixelhop), offers an unconditional generative model. Based on MNIST and Fashion-MNIST two datasets, GenHop can generate visually pleasant images whose FID scores are comparable with those of DL-based generative models.

In image forensics, steganography and steganalysis are like the two ends of the same coin. image steganography is a technique to conceal secret messages in the images by slightly modifying the pixel values. Corresponding to image steganography, steganalysis is the process to reveal the presence of the hidden message in images. Recently, steganalysis are focusing on defending content-adaptive steganographic schemes, for example WOW, HILL and S-UNIWARD, etc. Fig.1 [1] illustrates the modifications of cover image from different steganographic method.  Content-adaptive steganography is lean to do modifications on complex texture regions, which makes embedding traces less detectable for steganalyzers.

Traditionally, hand crafted features together with machine learning classifiers have good performance on steganalysis, such as Spatial Rich Model and its variants. After the emerging of neural networks, different CNN architectures are utilized in the steganalysis literature. Because of the important property that CNNs are able to extract complex statistical dependencies from high dimensional input and learn hierarchical representations, CNN-based features usually achieve better performance than traditional hand-crafted features. However, CNN based models are suffering from long training time, large model size and enormous consumption of computation resources.

We would like to utilize green learning methodology in steganalysis field, by incorporating Saab transform as feature extraction module in the future. Saab transform has shown its capability of extracting the high frequency representations in a feedforward way and preserving the light-weighted model size at the same time.

References:

[1] Tang, Weixuan, et al. “Adaptive steganalysis based on embedding probabilities of pixels.” IEEE Transactions on Information Forensics and Security 11.4 (2015): 734-745.

[2] C.-C. J. Kuo and Y. Chen, “On data-driven saak transform,” Journal of Visual Communication and Image Representation, vol. 50, pp. 237–246, 2018.

[3] C.-C. J. Kuo, M. Zhang, S. Li, J. [...]

Object detection is one of the most essential and challenging tasks in computer vision, while most state-of-the-art object detection methods adopt an end-to-end deep neural network, we aim at an interpretable framework that has low complexity, high efficiency in training, and high performance. The method is built upon the PixelHop framework, as shown in fig 1. The term “hop” denotes the neighborhood of a pixel. Pixelhop conducts spectral analysis on neighborhoods of different sizes centered on a pixel through a sequence of cascaded dimension reduction units. The neighborhoods of an object contain representative patterns of the objects such as salient contours and, as a result, they have distinctive spectral signatures at a certain scale that matches the object size, thus bounding boxes and  class labels can be predicted based on supervised learning with Saab coefficients in proper hops as the representations.

Our method takes YOLO’s problem formulation as reference and ensembles three major modules to finish the object detection task. As shown in fig.1, by proper settings of Pixelhop, we divide all the objects into three different scales, i.e. large(as shown blue), medium (as shown in green), and small (as shown in red), and have hops with proper receptive field (RF) responsible for proposing corresponding anchor boxes for different scales (as shown comparing with the “cat” example). With the Saab coefficients at each hop, we propose anchor boxes at each spatial location, and for each anchor box we train module 1 to predict its confidence score, module 2 to predict its class label, module 3 to predict its box regression. Eventually for each image our model will first propose potential boxes and use non max suppression based on confidence score to keep the best proposed [...]