NeurIPS 2021华为诺亚方舟实验室发表32篇主会议论文及3篇dataset track

NeurIPS 是全球最负盛名的 AI 学术会议，诺亚方舟实验室今年被收录35篇论文，去年同大会诺亚20篇。中稿方向涵盖AI基础理论、AI无损压缩、视觉、极简计算、transformer、强化学习、AutoML、数据集构建等。35篇中包含1篇Oral，2篇Spotlight论文和3篇dataset track，据大会官方统计，今年 NeurIPS 共有 9122 篇有效论文投稿，总体接收率 26%，只有 3% 被接收为 Spotlight 论文，Oral论文录取率低于1%。

我们将会对本次诺亚实验室研究工作进行多期系列专题报告。

1、强化学习

2、Out Of Distribution研究

3、AI无损压缩

4、优化算法理论

5、数据集构建

6、自动驾驶与基础模型

7、精简模型

Paper List：

Reinforcement Learning

1. Model-based reinforcement learning via imagination with derived memory.

2. A reinforcement learning based bi-level optimization framework for large-scale dynamic pickup and delivery problems.

3. An Efficient Transfer Learning Framework for Multiagent Reinforcement Learning.

4. Adaptive Online Packing-guided Search for POMDPs. 

5. Setting the Variance of Multi-Agent Policy Gradinets.

6. Discovering Multi-Agent Auto-Cirricula in Two Player Zero-Sum games.

7. Flattening Sharpness for Dynamic Gradient Projection Memory Benefits Continual Learning. 

Out-of-Distribution Generalization：

8. Towards a Theoretical Framework of Out-of-Distribution Generalization.

9. MixACM: Mixup-Based Robustness Transfer via Distillation of Activated Channel Maps.

10. No Fear of Heterogeneity: Classifier Calibration for Federated Learning with Non-IID Data.

AI Lossless Compression：

11. iFlow: Numerically Invertible Flows for Efficient Lossless Compression via a Uniform Coder. (Spotlight).

12. OSOA: One-Shot Online Adaptation of Deep Generative Models for Lossless Compression.

13. On the Out of Distribution Generalization of Probabilistic Image Modelling.

Optimization Theory：

14. Stability and Generalization of Bilevel Programming in Hyperparameter Optimization.

15. On Effective Scheduling of Model-based Reinforcement Learning.

16. Greedy and Random Quasi-Newton Methods with Faster Explicit Superlinear Convergence.

Dataset track:

17.  NATURE: Natural Auxiliary TextUtterances for Realistic Spoken Language Evaluation.

18.  SODA10M: A Large-Scale 2DSelf/Semi-Supervised Object Detection Dataset for Autonomous Driving.

19. One Million Scenes for Autonomous Driving: ONCE Dataset

Autonomous Driving and Basic Model：

20. Learning Transferable Features forPoint Cloud Detection via 3D Contrastive Co-training

21. Transformer in Transformer.

22. Augmented Shortcuts for VisionTransformers.

23. Neural Architecture Dilation for Adversarial Robustness

24. SOFT: Softmax-free Transformer with Linear Complexity (Spotlight).

25. Manifold Topology Divergence: a Framework for Comparing Data Manifolds.

26. Do NerualOptimal Transport Solvers Work? A Continuous Wasserstein-2 Benchmark

Efficient Model

27. Learning Frequency Domain Approximation for Binary Neural Networks (Oral).

28. Dynamic Resolution Network.

29. Post-Training Quantization for Vision Transformer.

30. Handling Long-tailed Feature Distribution in AdderNets

31. An Empirical Study of Adder Neural Networks for Object Detection

32. Adder Attention for Vision Transformer

33. Towards Stable and Robust AdderNets

34. S3 : Sign-Sparse-Shift Reparametrization for Effective Training of Low-bitShift Networks.

35. Demystifying and Generalizing Binary Connect. 

[关于转载]：本文转载于诺亚实验室 ，仅用于学术分享，有任何问题请与我们联系：report@aminer.cn。