| NGUYEN THE TUNG | D, 中間発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 吉野 幸一郎, Sakriani Sakti |
|
Title: Multimodal negotiation dialog system using end-to-end approach
Abstract: Negotiation dialog systems have great potential for many practical applications in daily life. However, current negotiation systems are built using pipeline approach, which makes adaptation to different domains problematic. In addition, most of the existing systems only use linguistic information for the dialog management process, while it was shown that other non-verbal information such as acoustic and visual are helpful for dialog management of negotiation systems. In this research, I propose an end-to-end model based on Deep Recurrent Q-Network, which takes input from linguistic, acoustic, and visual modalities and outputs the most apropriate system dialog action. The proposed model utilizes hierarchical tensor fusion to combine features from different modalities in an efficient way. In order to deal with the lack of large multimodal dialog corpus, I use Bayesian method to take uncertainty in the network's weight parameters, thus, increasing sample-efficiency when training. Language of the presentation: English | |||
| HAN XINYOU | M, 2回目発表 | 知能システム制御 | 杉本 謙二, 岡田 実, 松原 崇充, 小蔵 正輝 |
|
Title: SPR Establishment in Feedback Error Learning
Abstract: Feedback error learning (FEL) is an effective method for designing two-degree-of-freedom (2DOF) control system. In this method, stability is first ensured by a fixed feedback block, and then reference tracking is attained by a feedforward block tuned online. A tuning law is proposed under a certain strictly positive real (SPR) condition. However it is not easy to achieve the SPR condition. In my research, an approach to select the proper design parameter that achieves the SPR condition is proposed. This approach is based on a linear matrix inequality (LMI) and proved to be correct. Also, numerical simulations show its effectiveness. Language of the presentation: English | |||
| WILLIAMS FATOU | M, 2回目発表 | サイバーレジリエンス構成学 | 門林 雄基, 岡田 実, 安本 慶一, 妙中 雄三 |
|
Title: Enabling Large Data Transmission Over Low Power Wide Area Network(LPWAN)
Abstract: Low Power Wide Area Networks have brought a new dimension of innovation for smart city applications with its long range and low power capability. The integration of large data transmission such as images along with small payload size data in LoRa networks can contribute to realizing the goals of improved productivity and reduced management cost for applications. However, transmitting large data in multiple data frames is hindered by two main properties of LoRa: inefficient transmission scheduling due to random channel selection of ALOHA-like LoRa causing deadline misses and the unlicensed sub-GHz band duty-cycle regulation policy imposed on both nodes and channels that limits the transmission time of IoT devices. In this research, we proposed a duty-cycle aware priority-group-based link scheduling algorithm for delay constraint large data transmissions. Language of the presentation: English | |||
| MALLARI JUAN CARLO FLORES | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD. ALTAF-UL-AMIN. 黄 銘 |
|
Title: Ensembles of Time-Varying Models for the Inference of Gene Regulatory Networks
Abstract: Advancements in RNA sequencing technology and computational capability have allowed researchers to develop time-varying models for gene regulatory network (GRN) inference. Models for GRN inference, however, exhibit biases for particular network motifs, resulting in a lack of generalizability in their performance. We aim to address this problem by using model ensembles that allow for the elimination of individual model limitations. The performance with respect to changepoint identification and edge detection of an ensemble of three existing time-varying models (TESLA, EDISON, and HMDBN) is evaluated using life cycle data from Drosophila melanogaster and in vivo synthetic data from Saccharomyces cerevisiae. Future work involves the creation of a comprehensive metric for the evaluation of time-varying models and a novel approach to time-varying GRN inference that will be added to the ensemble. Language of the presentation: English | |||