ZhiruoZhou

Image anomaly detection and localization is a fundamental problem in pattern recognition and computer vision, with numerous applications in many areas, such as industrial manufacturing inspection, medical image diagnosis and even video surveillance analysis. The goal of image anomaly detection is to determine whether an input image contains an anomaly, and image anomaly localization is to locate the anomaly on the pixel level. Like most other anomaly detection problem, we formulate image anomaly detection as an unsupervised task, which means only normal images are available during model training. This is because anomalous examples are either too expensive or too few to model their distributions during the training, which also makes it an extremely challenging yet attracting problem.

To tackle this problem, we propose two method with both deep learning and successive subspace learning techniques.

We propose a new deep learning framework for unsupervised image anomaly detection and localization. Our model first utilizes an encoder to generate low-dimensional embeddings for local image patches, which are further fed into a density estimation network inspired by Gaussian Mixture Model. Given the low-dimensional patch embedding as input, density estimation network model the distribution of embedding like GMM clustering, and predict its cluster membership as output. Then, total probability of the give local patch could be computed and further used as a loss term to guide the learning process. Extensive experimental results show that the proposed method achieves very competitive performance compared with the state-of-the-art methods.
We are exploring to use successive subspace learning (SSL) to achieve a more efficient and interpretable method for image anomaly detection and localization. It first employees PixelHop++[1] as feature extractor, in which each hop could encode feature with different receptive field. Then, we [...]

In order to get an accurate perception of surrounding environment in different tasks including autonomous driving, robot navigation, and sensor-driven situational awareness, abundant environment information is necessary. This information can be obtained from different types of multimodal sensors, such as LiDAR sensors, electro-optical/infrared (EO/IR) cameras, GPS/IMU. Before using the collected data, information fusion among these sensors is a critical topic. Specifically, people want to utilize color and shape information from camera and distance information from LiDAR sensors. In which task, the process of finding correspondent points between two sensors is essential. This procedure is called multimodal sensor calibration, in which we need to find the 6DoF extrinsic parameters between these two sensors.

In this work, we develop a new deep learning-driven technique for accurate calibration of LiDAR-Camera pair, which is completely data-driven, does not require any specific calibration targets or hardware assistants, and the entire processing is end to end and fully automatic. We utilize the advanced deep neural network to align accurately the LiDAR point cloud to the image, and regress 6DoF extrinsic calibration parameters. Geometric supervision and transformation supervision are employed to guide the learning process to maximize the consistency of input images and point clouds. Given input LiDAR-Camera pairs as training dataset, the system automatically learns meaningful features, infers modal cross-correlations, and estimates the accurate 6DoF rigid body transformation between the 3D LiDAR and 2D image in real-time.

Images in slides show the system overview and experiment results. In experiment results, the background is the correspondent RGB image. The transparent colormap is the depth map, from blue to red corresponding small to a large distance. The first row is the input RGB images. The second row is input depth maps, the third row [...]

Let us hear what he has to say about his defense and an abstract of his thesis.

The success of machine learning algorithms depends heavily on data representation techniques. Natural language is one kind of unstructured data source which does not have a natural numerical representation as to its original form. In most natural language processing tasks, a distributed representation of text entities is necessary. In this dissertation, we investigate and propose representation learning techniques to learn embeddings for 1) words, 2) sentences and 3) knowledge graph entities.

We will first give an overview of our embedding research, including word embedding evaluation and enhancement, sentence embedding from both non-contextualized and contextualized word embeddings. Then, we will focus on two recent works: InductivE and DomainWE. InductivE is proposed to handle the inductive learning problem on commonsense knowledge graph completion task. The entity embeddings are directly computed from its textual descriptions and an iterative training framework is proposed to enhance unseen entities with more structural information. DomainWE aims to distill knowledge from large pre-trained language models and synthesis as static word embeddings. As an efficient and robust alternative to large pre-trained language models, DomainWE has a smaller model size which is particularly important for deployment purposes, yet demonstrates better performance comparing with generic word embeddings.

I would like to thank Professor Kuo who guides me step by step through the path from a fresh graduate to an academic researcher. Through the past years, I have learned a lot from Professor Kuo and our MCL members, which not only include academic skills but also life principles. One of the most important things for the Ph.D. study is to keep self-motivated and try your best to achieve your initial dreams. Finally, [...]

Our work on new MLP interpretation includes:

Interpretable MLP design [1]:

A closed-form solution exists in two-class linear discriminant analysis (LDA), which discriminates two Gaussian-distributed classes in a multi-dimensional feature space. In this work, we interpret the multilayer perceptron (MLP) as a generalization of a two-class LDA system so that it can handle an input composed by multiple Gaussian modalities belonging to multiple classes. Besides input layer lin and output layer lout, the MLP of interest consists of two intermediate layers,  l1 and l2.  We propose a feedforward design that has three stages: 1) from lin to l1: half-space partitionings accomplished by multiple parallel LDAs, 2) from l1 to l2: subspace isolation where one Gaussian modality is represented by one neuron, 3) from l2 to lout: class-wise subspace mergence, where each Gaussian modality is connected to its target class. Through this process, we present an automatic MLP design that can specify the network architecture (i.e., the layer number and the neuron number at a layer) and all filter weights in a feedforward one-pass fashion.  This design can be generalized to an arbitrary distribution by leveraging the Gaussian mixture model (GMM). Experiments are conducted to compare the performance of the traditional backpropagation-based MLP (BP-MLP) and the new feedforward MLP (FF-MLP).

MLP as a piecewise low-order polynomial approximator [2]:

The construction of a multilayer perceptron (MLP) as a piecewise low-order polynomial approximator using a signal processing approach is presented in this work. The constructed MLP contains one input, one intermediate and one output layers. Its construction includes the specification of neuron numbers and all filter weights. Through the construction, a one-to-one correspondence between the approximation of an MLP and that of a piecewise low-order polynomial is established. Comparison [...]

The success of deep learning and neural networks often comes at the price of a large number of labeled data. Weakly-supervised learning (WSL) is an important paradigm that leverages a large number of unlabeled data to address this limitation. The need for WSL has arisen in many machine learning problems and found wide applications in computer vision, natural language processing, and graph-based modeling, where getting labeled data is expensive and there exists a large amount of unlabeled data.

Among weakly-supervised graph learning methods, label propagation (LP) has demonstrated good adaptability, scalability, and efficiency for node classification. However, LP-based methods are limited in their capability of integrating multiple data modalities for effective learning. Due to the recent success of neural networks, there has been an effort of applying neural networks into graph-structured data. One pioneering technique, known as graph convolutional networks (GCNs), has achieved impressive node classification performance for citation networks. However, GCNs fail to exploit the label distribution in the graph structure and difficult to scale for large graphs.

In this work, we propose a scalable weakly-supervised node classification method on graph-structured data, called GraphHop, where the underlying graph contains attributes of all nodes but labels of few nodes. Our method is an iterative algorithm that overcomes the deficiencies in LP and GCNs. With proper initial label vector embeddings, each iteration contains two steps: 1) label aggregation and 2) label update. In Step 1, each node aggregates its neighbors’ label vectors obtained in the previous iteration. In Step 2, a new label vector is predicted for each node based on the label of the node itself and the aggregated label information obtained in Step 1. This iterative procedure exploits the neighborhood information and enables GraphHop to [...]