Changqin Quan

Alleviating adversarial attacks via convolutional autoencoder

In order to defend adversarial attacks in computer vision models, the conventional approach arises on actively incorporate such samples into the training datasets. Nonetheless, the manual production of adversarial samples is painful and labor intensive. Here we propose a novel generative model: Convolutional Autoencoder Model to add unsupervised adversarial training, i.e., the production of adversarial images from the encoded feature representation, on conventional supervised convolutional neural network training. To accomplish such objective, we first provide a novel statistical understanding of convolutional neural network to translate convolution and pooling computations equivalently as a hierarchy of encoders, and sampling tricks, respectively. Then, we derive our proposed Convolutional Autoencoder Model with the ‘adversarial decoders’ to automate the generation of adversarial samples. We validated our proposed Convolutional Autoencoder Model on MNIST dataset, and achieved the clear-cut performance improvement over the normal Convolutional Neural Network.

Long short-term memory networks for automatic generation of conversations

Human Machine Interface demands the communicative propriety that would be applied in various linguistic tasks. In this research, we develop an intelligent ‘chat bot’, which generates conversational sentences via recurrent neural network and its coupled memory unit, long short-term memory (LSTM). Word strings in conversations are considered as time series data. Using a single neural network model that performs a simple task of outputting the next word from the preceding word, a conversational sentence can be generated by connecting the words. In the experiment, we performed the linguistic ‘Turning Test’ to evaluate the proposed system.

Adaptive Generative Initialization in Transfer Learning

In spite of numerous researches on transfer learning, the consensus on the optimal method in transfer learning has not been reached. To render a unified theoretical understanding of transfer learning, we rephrase the crux of transfer learning as pursuing the optimal initialisation in facilitating the to-be-transferred task. Hence, to obtain an ideal initialisation, we propose a novel initialisation technique, i.e., adapted generative initialisation. Not limit to boost the task transfer, more importantly, the proposed initialisation can also bound the transfer benefits in defending the devastating negative transfer. At first stage in our proposed initialisation, the in-congruency between a task and its assigned learner (model) can be alleviated through feeding the knowledge of the target learner to train the source learner, whereas the later generative stage ensures the adapted initialisation can be properly produced to the target learner. The superiority of our proposed initialisation over conventional neural network based approaches was validated in our preliminary experiment on MNIST dataset.

Text mining and pattern clustering for relation extraction of breast cancer and related genes

With the number increase of biomedical literatures, biomedical relation extraction discovery from the literature represents a new challenge for researchers in recent years. Then, a system that automatically extracts the related genes to the targeted disease is required. In this paper, we explore text mining and pattern clustering for relation extraction of breast cancer and related genes. It can be considered an unsupervised method and labeled data is not necessary. We firstly extract the candidate genes related to breast cancer by checking the window distance between the appearance of genes and breast cancer in a sentence. Then, two different clustering approaches (simple clustering and K-means clustering) are applied for finding the candidate association words that indicate the relationship between breast cancer and genes. The comparison experiment demonstrates that simple clustering is superior to K-means clustering in this task.

Passive velocity field control of a redundant cable-driven robot with tension limitations

This paper proposes a novel dynamic control approach for a cable driven robot with high redundant actuation and cable tension limitations to perform tracking task while interacting with environment. In order for a cable-driven exoskeleton robot to execute the task smoothly and safely, it is necessary to consider the tracking motion performance as well as passivity when interacting with the environment under the conditions of the actuation cables redundancy and the pulling limitation. With the additional consideration of the maximum limitation of the cable tension, cable-driven robot actually can only apply a certain range of feasible wrench on the external environment, which makes the task executed by robot be restricted. In order to make designed wrench be feasible and keep the desired trajectory tracking ability, we present a new control method by extending passive velocity field control (PVFC) method considering tracking stability and passivity. The approach augmented a higher dimensional virtual flywheel dynamics in a specific orthogonal complement space of the cables actuation space. After the final adjustment of the designed wrench with respect to the cables constraint, this method is capable of driving the cable robot to complete the trajectory tracking task and realize the passivity.

Non-contact, real-time monitoring of heart rate with a Webcam with application during water-bed massage

Techniques for measuring physiological signals are gathering a lot of attention in recent years. In many cases, the techniques require to make direct contact between the physiological signal sensors with the subjects skin. Several studies were proposed to measure heart rate (HR) with standard web cameras under the condition that the subject is keeping the quiescent state. In this paper, we propose a new technique for measuring HR which allows the subject head movements during water-bed massage. For comparison purpose, HR was measured simultaneously using an electrocardiography (ECG) device during all sessions of the experiment. The result shows that our proposed technique is able to improve the accuracy of HR measurement and reduce the processing time to nearly one-third of the existing method. Our approach could be beneficial for applications that require real-time HR monitoring during subjects head is moving.

Estimation of an object's physical parameter by force sensors of a dual-arm robot

We are developing a Nursing-Care Robot in order to reduce the load in physical nursing care. The concept of this robot is to promote the cared persons by the robot to activate their own motion ability as long as possible. This may lead to the improvement of the cared persons movement volition and movement abilities. In order to realize safe and human friendly robot care tasks, full body manipulation is an important technology, for which it is necessary to estimate the subjects center of gravity from the contact positions and forces with the robots two arms. In this paper, we estimate the center of gravity of object based on the contact point and the contact force estimated by force sensor on both robot arms. The position of gravity center is important to realize care tasks stably. We performed experiments and simulations for the single point contact and dual points contact cases using a cylindrical object. As a result, it is found that although some errors were recognized in the experiments compared with the simulations, the relations between the contact positions and such errors were observed. Such experimental error mainly comes from the difference of shape between the real robot and the model of the robot in simulation. As the future work, we have to improve the robot model so as to get better estimated information.