コロキアムB発表

日時: 6月14日(火)3限(13:30-15:00)


会場: L3

司会: KIM Youngwoo
KUO CHENG-YU D, 中間発表 ロボットラーニング 松原 崇充, 和田 隆広, 佐々木 光

title: Probabilistic Model-based Reinforcement Learning for Robotics: Exploration Safety and High Complexity Dynamics

abstract: Compliant robots have attracted researchers’ attention due to their superior contact safety and high agility. However, capturing an analytical dynamics model for compliant robots is challenging but necessary for robot control. As an alternative solution, Reinforcement Learning (RL) has demonstrated its capability to learn complex tasks by trial and error. Nevertheless, compliant robots require on-site training due to the dynamics difference between the simulation and the real, where skill transfer approaches such as Sim2Real are not applicable. As an on-site RL approach, probabilistic Model-based RL (MBRL) has demonstrated its sample efficiency and is capable of learning robot dynamics for control planning. Even so, MBRL lacks a mechanism to prevent impacts during the learning process, and compliant robots only reduce the severeness of hits, damaging may occur. To reduce the risk during the MBRL learning process, we proposed an uncertainty-aware contact-safe MBRL that utilizes the robot's knowledge level to limit its action to alleviate the risk. As demonstrated through real robot experiments, the uncertainty-aware MBRL significantly reduces the impact's severeness during the MBRL process. Next, as a popular category of the compliant robot, Spring-loaded Inverted Pendulum-inspired Bipedal robots (SLIP-biped) has shown their high agility in locomotion skill in recent years. Nevertheless, SLIP-biped's high dimensionality and complexity require a high complexity MBRL to model. Although modeling a high complexity model is possible with a high-level dynamics model, the computation load of such a model may cause impractical control frequency. Therefore, exploring a simplified and compact dynamics model of SLIP-biped would be a key to increasing the feasibility of MBRL with real-time control. We proposed the Task-decomposed Energy-model-based RL (TD-EMBRL) to achieve a compact yet sufficient dynamics model for controlling SLIP-biped. Our approach's effectiveness is demonstrated by hopping skill acquisition with a simulated SLIP-biped configured to replicate a real SLIP-biped. We achieved hopping skills acquisition from scratch at a high control frequency.

language of the presentation: English

 
EDGAR ANAROSSI M, 2回目発表 ロボットラーニング 松原 崇充, 和田 隆広, 鶴峯 義久
title: *** Jagged Robotic Motion Generation by Utilizing a DMPs-based Framework ***
abstract: *** Robot automation is an important aspect within human civilization to realize improvements to our quality-of-life. Unfortunately, with the current development of robot automation, the usage of robots are is still mostly restricted to environments which is still heavily controlled such as factories to ensure the safety of humans and the robot itself. Robots need to be able to produce safe, accurate, and adjustable motions to enable the usage of robot automation in our everyday environment, features which are built into a trajectory generation method called Dynamic Movement Primitives (DMPs). Recent development in DMPs which utilizes a deep-learning framework allows the generation of DMPs parameters given any input that can be processed by a neural network. One issue that can be found within DMPs is its inability to reconstruct multiple stopping points within a trajectory. In this research, improvements to the previous deep-learning framework are explored with the goal of improving DMPs trajectory reconstruction. Improvements in trajectory accuracy and data efficiency can be seen in the early results produced by the proposed method compared to the previous DMPs based framework.***
language of the presentation: *** English ***
 
小寺 智仁 M, 1回目発表 ロボットラーニング 松原 崇充, 安本 慶一, 鶴峯 義久

title: *** Development of Deep Reinforcement Learning Algorithms Robust to Function Approximation Errors in Spiking Neural Networks *** 

abstract: *** In recent years, deep reinforcement learning has been introduced to edge robots, and there are high expectations for automated operation of edge robots. Since such edge robots are required for long time operation, it is important to reduce power consumption during computation.To this end, spiking neural networks are known as a low-power machine learning model that can be applied to neurochips, where NNs can be implemented with extremely low power consumption compared to conventional GPUs. However, spiking neural networks do not have a learning method that can achieve realistic performance for real problems. Therefore, one of the most promising learning methods for spiking neural networks is to convert a conventional neural network to a spiking neural network and operate it as a spiking neural network. However, the conversion causes function approximation errors, and in deep reinforcement learning, where regression accuracy is important, the problem is that the conversion reduces task accuracy. In this study, we propose a reinforcement learning algorithm method for spiking neural networks that can be trained with low bits and is robust to function approximation errors. *** 

language of the presentation: *** Japanese *** 

 

会場: L2

司会: Duong Quang THANG
古川 慧 M, 2回目発表 知能コミュニケーション 中村 哲, 渡辺 太郎, 作村 諭一
title:Applying Syntax–Prosody Mapping Hypothesis, Prosodic Well-Formedness Constraints, and Boundary-Driven Theory to Neural Sequence-to-Sequence Speech Synthesis
abstract:End-to-end text-to-speech synthesis (TTS), which generates speech sounds directly from strings of texts or phonemes, has improved the quality of speech synthesis over the conventional TTS. However, most previous studies have been evaluated based on subjective naturalness and have not objectively examined whether they can reproduce pitch patterns of phonological phenomena such as downstep, rhythmic boost, and initial lowering that reflect syntactic structures in Japanese. These phenomena can be linguistically explained by phonological constraints, boundary-driven theory, and the syntax–prosody mapping hypothesis (SPMH), which assumes projections from syntactic structures to phonological hierarchy. Although some experiments in psycholinguistics have verified the validity of the SPMH, it is crucial to investigate whether it can be implemented in TTS. To synthesize linguistic phenomena involving syntactic or phonological constraints, we propose a model using phonological symbols based on the SPMH, boundary-driven theory, and prosodic well-formedness constraints. Experimental results showed that the proposed method synthesized similar pitch patterns to those reported in linguistics experiments for the phenomena of initial lowering and rhythmic boost. The proposed model efficiently synthesizes phonological phenomena in the test data that were not explicitly included in the training data.
language of the presentation: Japanese
 
QI HELI M, 1回目発表 知能コミュニケーション 中村 哲, 渡辺 太郎

title: *** The research on semi-supervised ASR based on machine speech chain and pseudo-labeling techniques *** 

abstract: *** Thanks to the advancement of deep neural networks, sequence-to-sequence (S2S) automatic speech recognition (ASR) has made significant progress. S2S ASR models are designed for directly converting the input speech into transcripts. However, a large amount of labeled speech-text data is essential for training S2S ASR models to achieve state-of-the-art performance. In recent years, there has been more and more research on semi-supervised learning algorithms for efficiently training S2S ASR models using as less labeled data as possible with the aid of numerous unlabeled data. Among them, pseudo-labeling is a simple but efficient technique to help ASR models learn more from untranscribed speech. In this research, I plan to combine this technique with the machine speech chain to lift the limit of semi-supervised ASR. In my first attempt, I utilized the machine speech chain as a data augmentation method and applied it with consistency regularization, a pseudo-labeling algorithm. My experimental results show that the performance of supervised baseline ASR models can be significantly improved with the help of untranscribed speech. *** 

language of the presentation: *** English *** 

 
LIU CHANG M, 1回目発表 自然言語処理学 渡辺 太郎, 中村 哲, 進藤 裕之

title: *** Improving Cross-lingual Sentiment Analysis with Substituting Words Approach on Multilingual Pre-training Models *** 

abstract: *** Sentiment Analysis are well-studied in mono-lingual settings, particularly in English. Since Sentiment Analysis need annotated language datasets. In order to perform Sentiment Analysis studies in languages that lack annotated datasets. Choose cross-lingual projections through bilingual dictionaries. Alternatively, simply apply machine translation as a Sentiment Analysis pre-processing step. However, difference in word order between source and target languages is a problem for cross-lingual models that use bilingual embeddings as features. Code-switching approach has good generalization capability in multi-language field. With the help of multilingual-BERT, we can achieve sentiment analysis by transforming arbitrary languages into code-mixed sentences with different languages through word substitution. It is a solution for low resource languages that do not have enough training datasets. *** 

language of the presentation: *** English *** 

 
的川 雄飛 M, 1回目発表 自然言語処理学 渡辺 太郎, 中村 哲, 進藤 裕之
title: *** Identification of Synonymity of Synonymous Katakana Words Using Comparison to Foreign Notations ***
abstract: *** To identify synonymity of synonymous katakana words, I plan to convert katakana words and their corresponding foreign language notations to IPA (International Phonetic Alphabet) and calculate the similarity. Using a dataset for transliteration, which converts notations written in one writing system to another type of characters, I plan to measure the performance of the model using IPA in the task of outputting the correct ID of a named entity from a katakana word as input. ***
language of the presentation: *** Japanese ***
発表題目: *** 外国語表記との比較を用いた同義カタカナ語の同義性特定 ***
発表概要: *** 同義カタカナ語の同義性を特定するため、カタカナ語とそれに対応する外国語表記をそれぞれ国際音声記号に変換し、類似度を計算する。 ある文字体系を用いて書かれた表記を別の種類の文字に変換する翻字のためのデータセットを利用し、カタカナ語を入力として正しい固有表現IDを出力するタスクにおいて国際音声記号を用いたモデルの性能を測る。 ***
 
VINCENT MICHAEL SUTANTO M, 1回目発表 自然言語処理学 渡辺 太郎, 荒牧 英治, 進藤 裕之

title: Transformers-based Image to Latex Equation Conversion

abstract: Image to LaTex equation is a task proposed by OpenAI that aims to build an End-to-end Neural Markup Generation that can receive an image of the equation and generate its respective LaTeX markup. Current research has used CNN as the encoder and RNN as the decoder. However, deep CNN and RNN layers are needed to capture global dependencies within the images and the markup tokens. A transformers-based model allows us to alleviate this problem, as the attention module inside can capture global dependencies from the start. Therefore, this research proposed a full transformers encoder-decoder architecture to derive LaTeX markup from its respective image equation.

language of the presentation: English