News

Let us hear what she wants to say about her defense and an abstract of her thesis.

Deep learning techniques utilize networks with multiple layers cascaded to map the inputs to desired outputs. To map the entire inputs to desired outputs, useful information should be extracted through the layers. During the mapping, feature extraction and prediction are jointly performed. We do not have direct control for feature extraction. Consequently, some useful information, especially local information, is also discarded in the process.

In this thesis, we specifically study local-aware deep learning techniques from four different aspects: 1) Local-aware network architecture 2) Local-aware proposal generation 3) Local-aware region analysis 4) Local-aware supervision

Specifically, we design a multi-modal attention mechanism for generative visual dialogue system, which simultaneously attends to multi-modal inputs and utilizes extracted local information to generate dialogue responses. We propose a proposal network for fast face detection system for mobile devices, which detects salient facial parts and uses them as local cues for detection of entire faces. We extract representative fashion features by analyzing local regions, which contain local fashion details of humans’ interests. We develop a fashion outfit compatibility learning method, which models each outfit as a graph and learns outfit compatibility using both global and local supervisions on the graphs.

I would like to thank Prof. Kuo and all the lab members for their help. I have learned a lot through my PhD journey and I want to share some feelings and experiences. One essential part of the doctoral training is mental training, from which I have become more persistent, self-disciplined and motivated. As this journey may take several years, maintaining a balanced life is very important. I wish the best to all the lab members and [...]

In the task of fashion compatibility prediction, the goal is to pick an item from a candidate list to complement a partial outfit in the most appealing manner. Existing fashion compatibility recommendation work comprehends clothing images in a single metric space and lacks detailed understanding of users’ preferences in different contexts. To address this problem, we propose a novel Metric-Aware Explainable Graph Network (MAEG). In MAEG, we leverage a Latent Semantic Extraction Network (LSEN) to obtain representations of items in the metric-aware latent semantic space. Then, we develop a graph filtering network and Pairwise Preference Attention (PPA) module to model the interactions between users’ preferences and contextual information. With MAEG, we can provide recommendation to users as well as explain how each item and factor contribute to the final prediction. Extensive experiments on two large-scale real-world datasets reveal that MAEG not only outperforms the state-of-the-art methods, but also provides interpretable insights by highlighting the role of semantic attributes and contextual relationships among items.

MCL Director, Professor Kuo, gave an invited speech at Kyoto University on December 19, 2019. The title of his speech was “From Feedforward-Designed Convolutional Neural Networks (FF-CNNs) to Successive Subspace Learning (SSL)”.  Professor Kuo’s visit to Kyoto University was hosted by Professor Tatsuya Kawahara. The lecture was also an event of IEEE SPS Kansai Chapter.

The abstract of his speech is given below. “Given a convolutional neural network (CNN) architecture, its network parameters are typically determined by backpropagation (BP). The underlying mechanism remains to be a black-box after a large amount of theoretical investigation. In this talk, I will first describe a new interpretable feedforward (FF) design with the LeNet-5 as an example. The FF-designed CNN is a data-centric approach that derives network parameters based on training data statistics layer by layer in a one-pass feedforward manner. To build the convolutional layers, we develop a new signal transform, called the Saab (Subspace approximation with adjusted bias) transform. The bias in filter weights is chosen to annihilate nonlinearity of the activation function. To build the fully-connected (FC) layers, we adopt a label-guided linear least squared regression (LSR) method. To generalize the FF design idea furthermore, we present the notion of “successive subspace learning (SSL)” and present a couple of concrete methods for image and point cloud classification. Experimental results are given to demonstrate the competitive performance of the SSL-based systems. Similarities and differences between SSL and deep learning (DL) are compared.”

MCL Director, Professor Kuo, was invited by Korean National IT Industry Promotion Agency (NIPA) to deliver a keynote speech at the Grand Cloud Conference, Seoul, Korea on Dec. 3, 2019. The title of Professor Kuo’s talk was “AI Cloud Strategy for Government and Industry”. Here is the abstract.

There have been many successful artificial intelligence (AI) stories in the last decade due to big data analytics and the resurgence of neural-network-based deep learning. For clarity, I would like to give my own definition about AI. That is, AI is not about competing with humans in terms of intelligence but an advanced form of “automation”, which leverages both data and domain knowledge to achieve domain-specific tasks more effectively. Then, I will discuss the AI deployment strategy. It is centered around “smart environments” such as smart homes, smart cities, etc. We can leverage sensors, Internet of Things (IoT) and 5G infrastructures to acquire a large amount of data for better decision making. After that, I will mention three R&D opportunities: 1) explainable AI, 2) integrated rule-based/data-driven methodology and 3) federated learning. For the first topic, AI is not necessarily dependent upon deep learning, which is a black box. We are working on an alternative approach, called Successive Subspace Learning (SSL), to solve the big data problem at USC. SSL is mathematically transparent and expandable depending on the training data sizes and classes. It has great potential. For the second topic, I want to emphasize that big data is helpful but not sufficient by itself alone. It is more powerful to have an integration of the “rule-based” and the “data-driven” approaches. The integration allows better generalizability and robustness. The third topic is concerned with data-driven model training without [...]

Dr. Sachin Chachada, a former MCL alumnus, and Professor C.-C. Jay Kuo received the 2019 Sadaoki Furui Paper Award at the open ceremony of the 2019 APSIPA ASC held in Lanzhou, China, on November 19 for their paper below:
Sachin Chachada and C.-C. Jay Kuo, “Environmental sound recognition: a survey,” Published online: 15 December 2014, e14, APSIPA Trans. on Signal and Information Processing.
The paper has been cited by 107 in Google Scholar, the number of downloads in 2019 (to end of Sept.) is 1523. The abstract of the paper is given below.
Although research in audio recognition has traditionally focused on speech and music signals, the problem of environmental sound recognition (ESR) has received more attention in recent years. Research on ESR has significantly increased in the past decade. Recent work has focused on the appraisal of non-stationary aspects of environmental sounds, and several new features predicated on non-stationary characteristics have been proposed. These features strive to maximize their information content pertaining to signal’s temporal and spectral characteristics. Furthermore, sequential learning methods have been used to capture the long-term variation of environmental sounds. In this survey, we will offer a qualitative and elucidatory survey on recent developments. It includes four parts: (i) basic environmental sound-processing schemes, (ii) stationary ESR techniques, (iii) non-stationary ESR techniques, and (iv) performance comparison of selected methods. Finally, concluding remarks and future research and development trends in the ESR field will be given.