| OCHOA REY CYNTHIA MARIA | D, 中間発表 | ロボットラーニング | 松原 崇充, 和田 隆広, 鶴峯 義久 |
|
title: Interactive Imitation Learning from Demonstrations with Symbolic Goals Assigning for Long Horizon Robotic Task
abstract: Interactive Imitation Learning (IIL) is a method where the robot learns how to perform a task from human demonstrations provided only in unknown states. Through IIL, tedious and expert programming are avoided while keeping data collection efficiency and robustness. Despite the former advantages, IIL has not been applied to hierarchical control architecture, which has been proven its suitability to tackle long horizon tasks diving it into symbolic sub-goals. Therefore, this research aims to propose a framework for Interactive Imitation Learning from human demonstrations applied to a hierarchical architecture. The effectiveness of this proposal is verified by simulation of cooking task in a kitchen environment. The results show the potential use of this approach for more complex and longer tasks. language of the presentation: English | |||
| BERNARD TIONG ING SHENG | M, 1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, 佐々木 光 |
|
Title: Multi-agent Reinforcement Learning with Intention Sharing For Physical Robotics Collaborative Tasks
Abstract:Reinforcement learning (RL) has achieved remarkable success in various complex control problems such as robotics. Multi-agent reinforcement learning (MARL) extends RL to multi-agent systems, which model many practical real-world problems involving multiple agents, such as coordinating robots to perform collaborative tasks. However, physical robotics collaborative tasks require strong coordination and synchronization skills are still cannot be solved smoothly and completely by conventional MARL. Such complicated tasks requiring strong interaction between agents so far haven't been addressed in the MARL field of study. The primary way for humans to interact with each other is through communication. Unlike the conventional method used in MARL communication, where agents share their own observations/actions, humans also share their intended actions/future plans to complete the collaborative tasks smoothly. Rather than raw observations, we intend to use agents' intentions as communication messages to enhance their coordinated behavior. language of the presentation: English | |||
| RAMOS FERNANDEZ ALONSO | M, 1回目発表 | ロボットラーニング | 松原 崇充, 池田 和司, 鶴峯 義久, 佐々木 光 |
|
title: Quantized P-Network Distillation (QPD)
abstract: Edge robots, unlike installation robots, have application on a large area: crime prevention, facility inspection, transport, etc. Due to this, they have limited power source (limited operation time). Neuromorphic chips (edge-controllers) are recently attracting a lot of attention for edge-robots due to its low power consumption. However, using neuromorphic chips carries the use of Spiking Neural Networks (SNNs), and therefore some difficulties: SNNs are difficult to learn and have low approximation accuracy. The first solution is Quantized P Network (QPN): based on Deep Q Network (DQN), QPN uses Quantized Neural Network (QNN) to be robust to SNN function approximation errors. However, since for long horizon task high function approximation accuracy is required during learning, QPN does not work for this kind of task. To overcome this problem, we propose Quantized P-Network Distillation (QPD); with QPD we learn a Floating-Point Neural Network (FPNN), with high approximation accuracy and then transfer this learning to a QNN. The results show that QPD can overcome QPN for long horizon tasks. language of the presentation: English | |||
| LU JINZHENG | M, 2回目発表 | 光メディアインタフェース | 向川 康博, 中村 哲, 舩冨 卓哉, 品川 政太朗, 藤村 友貴, 北野 和哉 |
|
title: Improving the Temporal Consistency for Automatic Anime Colorization
abstract: Automatic coloration is a crucial task in animation production. Traditional approaches have typically processed each frame individually to color the line draft, but this method has limitations in improving the efficiency of animation production. In this study, we propose a novel approach utilizing the TPS (Temporal Pseudo Supervision) technique to process consecutive frames of animation videos, enhancing temporal consistency and segmenting the automatic coloring areas using a semantic segmentation model. This approach leverages the pseudo labels obtained from the semantic segmentation of adjacent frames to optimize the coloring effect, thereby improving the temporal consistency within the video. language of the presentation: English | |||
| SARHANGZADEH ARMIN | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, 中村 哲, 進藤 裕之 |
|
title: Towards Low-Latency Neural IMEs for Japanese
abstract: Japanese input method editors (IMEs) are essential tools for inputting Japanese text using a limited set of characters such as the kana syllabary. However, despite their importance, the potential of newer attention-based encoder-decoder neural networks, such as Transformers, has not been fully explored for IMEs due to their high computational cost and low-quality intermediate output in simultaneous settings, leading to high latencies. We thus propose a simple decoding policy for low-latency simultaneous kana-kanji conversion, the main component of Japanese IMEs, leveraging the monotonic nature of the process to achieve incremental anticipation-free conversion. language of the presentation: English | |||
| LIU CHANG | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, 中村 哲, 進藤 裕之 |
|
title: Code-Switching Method for Low-Resource Language Sentiment Analysis on Multilingual Pre-training Model
abstract: Sentiment analysis on low-resource languages can be limited by the lack of a sufficient corpus to build a high-performance model. Parallel corpus and machine translation approaches do not provide sufficient access to cross-lingual information for sentiment analysis. We propose a code-switching method based on multilingual embeddings and Multilingual BERT (mBERT) masked language modeling that generates artificial code-switching sentences. This data augmentation is then used to fine-tune the pre-trained multilingual model. The experiment results show that our approach effectively adapts the embedding of high-resource languages to low-resource target languages. language of the presentation: English | |||