ICLR2018 Workshopに当研究室の論文が5件採択されました。
Neuron as an Agent
【Information】
Shohei Ohsawa, Kei Akuzawa, Tatsuya Matsushima, Gustavo Bezerra, Yusuke Iwasawa, Hiroshi Kajino, Seiya Takenaka, Yutaka Matsuo: Neuron as an Agent, International Conference of Learning Representation (ICLR18) Workshop, 2018
【Overview】
Designing Efficient Neural Attention Systems Towards Achieving Human-level Sharp Vision
【Information】
A.R. A.Ghani, N. Koganti, A. Solano, Y. Iwasawa, K. Nakayama, Y. Matsuo: Designing Efficient Neural Attention Systems Towards Achieving Human-level Sharp Vision, International Conference of Learning Representation (ICLR18) Workshop, 2018
【Overview】
We investigate the effect of using a hierarchy of visual streams in training an efficient attention model towards achieving a human-level sharp vision. We perform our evaluation on a simulated human-robot interaction task where the agent attends to faces that are looking at it.
Figure: Robot and simulator setups, we used two cameras (with wide and focused lenses), and generated four hierarchical streams as input for the network, the FOV of each stream is been selected to help in HRI task. where the widest stream with 120◦ can locate the person to attend to, the stream with 30◦ can locate parts of the human body such as a face, the stream with 10◦ can direct the sharpest stream to the region of interests for inference ROI.
Expert-based reward function training: the novel method to train sequence generators
【Information】
Joji Toyama, Yusuke Iwasawa, Kotaro Nakayama, Yutaka Matsuo: Expert-based reward function training: the novel method to train sequence generators, International Conference of Learning Representation (ICLR18) Workshop, 2018
【Overview】
系列生成器の訓練方法として,SeqGANに代表される,敵対的学習と方策勾配法を組み合わせた方法が知られているが,本研究では敵対的学習を代替する訓練方法を提案する.
提案手法では,生成器は提案分布と真の分布の尤度比を報酬とみて方策勾配法で学習することで,安定的な学習を達成する.
Extending Robust Adversarial Reinforcement Learning Considering Adaptation and Diversity
【Information】
Hiroaki Shioya, Yusuke Iwasawa, Yutaka Matsuo: Extending Robust Adversarial Reinforcement Learning Considering Adaptation and Diversity, International Conference of Learning Representation (ICLR18) Workshop, 2018
【Overview】
強化学習において敵対的訓練を用いて環境の違いに対するロバスト性を向上させるRobust Adversarial Reinforcement Learningの拡張を2つ提案しました。テスト環境との違いを制約に用いる拡張と、敵対的訓練に複数の敵対者を用いる拡張によって従来法と比較して性能を高めました。
Censoring Representations with Multiple-Adversaries over Random Subspaces
【Information】
Yusuke Iwasawa, Kotaro Nakayama, Yutaka Matsuo: Censoring Representations with Multiple-Adversaries over Random Subspaces, International Conference of Learning Representation (ICLR18) Workshop, 2018