NetEase Fuxi has 3 papers selected in ICLR 2023, covering areas such as reinforcement learning and natural language processing.

The 11th International Conference on Representation Learning (ICLR) is expected to be held offline in Kigali, the capital of Rwanda, from May 1st to 5th. Recently, ICLR announced the paper acceptance results, including a total of 3 papers by NetEase Fuxi. Among these three papers, one was selected as an oral presentation paper and the other two were selected as spotlight presentation papers. The content of these papers involves many fields such as reinforcement learning and natural language processing. The paper selected this time is an important achievement of NetEase Fuxi team in these research directions, and it is also their recognition and outstanding contribution in the academic world.

Experiments show that KLD is more sensitive to abnormal points, while TCD is robust.

In order to balance the estimation of TVD, we introduce the TaiLr target. TaiLr achieves this goal by reducing the weight of real data samples with low model probability, and the penalty strength can be adjusted as needed. Experiments demonstrate that our method mitigates overestimation of degenerate sequences while maintaining diversity, and improves the generation quality for a wide range of text generation tasks.

However, past work often focuses on pre-training a strategy with different skills through exploring the environment. However, it is difficult to ensure the performance improvement of downstream tasks through pre-training methods of diversified exploration, and may even lead to The greater the pre-training consumption, the lower the performance will be due to the "mismatch" problem. Therefore, NetEase Fuxi and the Tianjin University Deep Reinforcement Learning Laboratory team proposed the EUCLID framework, which introduces a model-based RL paradigm to benefit from accurate dynamic models through long-term pre-training to achieve rapid downstream task adaptation and Higher sampling efficiency. In the fine-tuning phase, EUCLID uses pre-trained dynamic models for policy-guided planning. This setting can eliminate performance shocks caused by mismatch problems and obtain monotonous performance improvements.

The experimental results show that NECSA achieved the highest scores in all experimental environments and reached the state-of-the-art level.

NECSA can be easily integrated into reinforcement learning algorithms and has strong versatility. One of the typical application scenarios is the training of game competition robots. NECSA provides a new idea based on state analysis, which can enhance the learning effect and is especially suitable for complex and high-dimensional game state representation. Through NECSA, the competitive level and anthropomorphism of the robot can be optimized better and faster, and good model interpretability can be provided. In the future, NetEase Fuxi will promote the practical application of the NECSA method in multiple game scenarios.

Special thanks to the team of Professor Huang Minlie of Tsinghua University for their important research contribution to "Tailoring Language Generation Models under Total Variation Distance". Their research work has made important contributions in the customization of language generation models, providing new ideas and methods for improving natural language processing technology. At the same time, we would like to thank the Deep Reinforcement Learning Laboratory of Tianjin University for its important research contribution to "EUCLID: Towards Efficient Unsupervised Reinforcement Learning with Multi-choice Dynamics Model". Their research work focuses on the field of unsupervised reinforcement learning, and proposes an efficient multi-choice dynamic model, making important contributions to the development of reinforcement learning algorithms. In addition, we would also like to thank Kyushu University Pangu Laboratory for its important research contribution to "Neural Episodic Control with State Abstraction". Their research work focuses on neuron memory control and state abstraction, and proposes a novel neuron control method, which provides new ideas and technical support for the development and application of intelligent systems. The contributions of these research teams are not only important in academia but also have potential implications for practical applications. We express our sincere gratitude to them for their outstanding work and look forward to their continued success in their respective fields. As the top domestic game and pan-entertainment AI research and application institution, NetEase Fuxi is committed to opening up AI technology and products to more people. Multiple partners to promote the application of artificial intelligence technology in various fields. So far, more than 200 customers have chosen NetEase Fuxi's services, and the number of calls has exceeded hundreds of millions every day.

The above is the detailed content of NetEase Fuxi has 3 papers selected in ICLR 2023, covering areas such as reinforcement learning and natural language processing.. For more information, please follow other related articles on the PHP Chinese website!