| 石原 実 | M, 2回目発表 | 光メディアインタフェース | 向川 康博 | 松原 崇充 | 舩冨 卓哉 | 藤村 友貴 | 北野 和哉 | |
Title: Representation of Geometric Transform Field using Neural NetworkAbstract: Abstract Conventional methods for describing geometric transform fields using kernel regression have faced the challenge of computational complexity that scales with the number of sample points. In this study, we leverage neural networks to achieve regression where computational complexity remains constant regardless of the increase in sample points. We validated our approach on a registration task using human embryo slice images, confirming the feasibility of describing geometric transform fields with neural networks. Future work includes verifying the computational advantages of our method on larger datasets. Language of the presentation: Japanease発表題目: ニューラルネットワークによる幾何変換場の記述
発表概要: 従来のカーネル回帰による幾何変換場の記述法は、
サンプル点の増加に対して計算量が膨大になるという問題があった。 | ||||||||
| 福田 竜平 | M, 2回目発表 | ロボットラーニング | 松原 崇充 | 安本 慶一 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 | |
|
title: Imitation learning for long-term tasks using skill decomposition
abstract: Automation of long-term tasks by robots is expected. To automate long-term tasks, an approach in which a robot learns to imitate actions demonstrated by a human is essential. However, the demonstration of a long-term task involves complex actions composed of multiple skills, which makes the learning process difficult. This study proposes imitation learning that decomposes a demonstration into individual skills. The performance of the method is verified by decomposing the demonstration based on specific skill conditions in the pick-and-place task of an object. language of the presentation: Japanese | ||||||||
| 本間 天譲 | M, 2回目発表 | ロボットラーニング | 松原 崇充 | 安本 慶一 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 | 角川 勇貴 |
|
title: Sim-to-Real Reinforcement Learning for Complex Tasks on Neurochip-Driven Edge Robots
abstract: Neurochips are a computing device suitable for function approximation of spiking neural networks (SNNs), and because of their energy-saving and real-time performance, they have attracted attention as a computing device for control policies in autonomous control tasks of edge robots. Sim-to-real reinforcement learning is useful as a method for acquiring control policies for edge robots, but because SNNs approximate functions with a set of spike signals, learning with domain randomization is difficult due to the accuracy of the approximation. In previous studies on robot control policy acquisition using SNNs, methods have been developed to learn SNNs using gap amplification and distillation, but the instability of learning in domain randomization remains an issue, and Sim-to-Real has not been achieved. In this study, therefore, to solve this issue, we propose an SNN policy update method that combines gap amplification and distillation to strengthen the approximation ability of SNNs and improve the stability and accuracy of policy updates. The contributions of this study are to realize Sim-to-Real using SNNs for the first time, propose a new learning method that combines gap amplification and distillation, and demonstrate its effectiveness. As an experiment, we verified the effectiveness of this method in a maze task, and the results showed that the proposed method outperformed conventional methods, confirming its effectiveness in sim-to-real tasks. language of the presentation: Japanese | ||||||||