20201027

In recent years, security cameras and smartphones image have been used for criminal investigations.

Conventionally, humans visually collated images with suspects.

CNN is useful for estimating physical characteristics, but there is a trade-off between estimation accuracy and robustness.

In this seminar, we will propose a suitable CNN that has both estimation accuracy and robustness.

【論文紹介】タスクに依らない転移学習の統一的なアプローチが可能なモデルT5を提案し，併せて新しいデータセットC4を開発した。T5は様々な自然言語処理タスクをText-to-Textという1つの統合フレームで解く構造をとっており，近年SOTAを達成した多数のモデルの様々なアプローチを比較可能にした。様々なアプローチの包括的調査により，何が効果的かを判断し，その結果をモデルに反映させることで多くのタスクでSOTAを達成した。

Reinforcement Learning is attracting attention as the control system for driving robot in fuel consumption measurement. However, it is necessary to improve the sample efficient for practical use. In this research, we aim to get the transferable policy in driving task. In this seminar, I will show the result of experiment about MCP with prioritized sampling.