日時: 7月13日（水）3限 (13:30-15:00)

Title: Realization of Compressed Sensing Based Channel Estimation for OFDM System

Abstract: Compressed sensing (CS) is one of the hottest research topics in the sparse signal reconstruction problem. But CS implementation has a drawback of high computational complexity due to calculation between large size of matrices. We will propose a low-complexity CS hardware realization for channel estimation in orthogonal frequency division and multiplexing (OFDM) system using several optimization methods to reduce the implementation complexity of CS usage. Since the measurement matrix of CS computation is a truncated discrete Fourier transform (DFT) matrix, We can exploit the symmetrical property of this DFT matrix to significantly reduce its multiplication complexity and random access memory (RAM) usage. To achieve fast reconstruction period, we also provides a hardware architecture for the proposed method and its realization in field programmable gate array (FPGA). The simulation results show that the proposed methods can achieve lower complexity CS based channel estimation with almost the identical system performance with the conventional method. Moreover, the realized hardware can achieve the fastest execution time compare to that of other existing methods.

Language of the presentation: English

Title: Development and Mining of a Volatile Organic Compound Database

Abstract: Volatile organic compounds (VOCs) are small molecules that exhibit high vapor pressure under ambient conditions and have low boiling points. VOCs are produced naturally by living organisms and have important roles in chemical ecology and human health. Information on VOCs is scattered in the literature until now, however, there is still no available database describing VOCs and their biological activities. In this study, I developed a novel VOC database of microorganisms, fungi, and plants as well as human beings, which comprises the relation between emitting species, the volatiles, and their biological activities. I deposited the VOC data into KNApSAcK Metabolite Ecology Database. This database can be accessed online. Apart from the database development, I also analyzed the VOC data using hierarchical clustering and network clustering based on DPClus for the classification of VOC emitting species. In addition, I also compared several clustering methods based on Tanimoto coefficient as the similarity index of the chemical structure to cluster all VOCs emitted by various biological species to understand the relationships between chemical structures of VOCs and their biological activities. My research finding indicates that similar chemical structures of VOCs indicate possibilities of exhibiting similar biological activities. In future work, I will apply supervised machine learning methods such as deep neural network (DNN) and support vector machine (SVM) to develop classification models for predicting the biological activities of VOCs based on their structures.

Language of the presentation: English

Title: A Reinforcement Learning Model of the Shepherding Task

Abstract: Herding behavior is the collective unwilling behavior of a group being led by individuals to move in a single direction to a specified target. In this work, we will focus on a specific kind of herding behavior found in a flock of sheep being led by a dog, which we will call as the shepherding task. A heuristic model of the shepherding task was developed by Strombom, et al, where interaction rules between the dog and sheep, and among individual sheep are identified. Although the model explains the behaviors of the dog and sheep, how they learn the behaviors is not clear. Thus, we propose to create a reinforcement learning model of the shepherding task based on the existing heuristic model by Strombom, et al. This study reconstructs the shepherding task using SARSA, an algorithm for learning the optimal policy in reinforcement learning. Results show that with a discretized state and action space, the dog is able to successfully herd a flock of a sheep to the target position by first learning to reach a subgoal. A reward is given when the dog reaches the neighbourhood of a subgoal, while a penalty is incurred for each time the shepherding task is not completed. Finally, we present an example of a completed shepherding task which shows the agent's continuous success after the 350th episode.

Language of presentation: English

abstract: Lexical Simplification is the process of changing difficult words in a sentence to easier ones of the same meaning without modifying the meaning of the original sentence. We apply Lexical Simplification on Japanese text to help second language learners understand the original Japanese text. Unlike for English, for Japanese, there is no parallel corpus of easy and difficult texts publicly available. However, there are some free Japanese texts, such as BCCWJ, Mainichi Newspaper, vocabulary lists for JLPT and essays written by children. We propose a method to find pairs of easy and difficult words using this publicly available data. Our method is to firstly calculate word similarity using word embeddings. Then, we measure the difficulty level of similar words using frequency, technical word classification, JLPT vocabulary lists, and vocabularies from children's corpora.