Intelligence systems

[NirViaje]

Frequency principle: Fourier analysis sheds light on deep neural networks

• F-Principle：初探理解深度学习不能做什么

[NirViaje]

DeepMind on StarCraft

Reddit AmA

• https://www.reddit.com/r/MachineLearning/comments/ajgzoc/we_are_oriol_vinyals_and_david_silver_from/
  • https://mp.weixin.qq.com/s/axr5VFbHQmYo0shW9ilBaQ

• https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/
  • https://mp.weixin.qq.com/s/EHaFurtFhX8F8hyRYWmWew

为了支持大批不同版本 AlphaStar 智能体的对战与更新，DeepMind 构建了一个大规模可拓展的分布式训练环境，其中使用了最新的谷歌 TPUv3，这个训练环境可以支持成群的 AlphaStar 智能体实例同时运行；星际 2 游戏主体也有数千个实例同步运行。AlphaStar league 的自我对战训练过程用了 14 天，每个 AlphaStar 智能体使用了 16 个 TPU，最终相当于每个智能体都有长达 200 年的游戏时间。训练结束后的模型在单块消费级 GPU 上就可以运行。

虽然这次的模型用在了星际争霸 2 AI 上，但 DeepMind 认为这是个足够复杂、具有代表性的任务，用来解决这个任务的技术也可以用在更多其他的复杂问题上。比如这个为长序列建模设计的网络架构可以用在更多不完全信息的长序列建模任务中，比如天气预测、气候建模、语言理解等等。他们也会继续开发 AlphaStar 项目，并利用其中的技术收获改善更多的任务。

另一方面，DeepMind 认为此次设计的训练策略也是通往安全、鲁棒的 AI 的一条新路径。现阶段的 AI 系统的一大难题就是难以预测系统会在多少种不同的情况下发生失效，星际争霸的人类职业选手们取胜 AI 也往往就是依靠寻找并攻击 AI 的弱点和错误。AlphaStar 中提出的群体训练策略就是一种可靠得多、出错的可能性明显减小的训练策略。DeepMind 的研究人员们认为这种方法还有很多潜力可以挖掘，也许未来它就会成为安全因素非常关键的问题中的重要一环。而最终，DeepMind 希望可以创建出真正智慧的系统，帮助人类解决一些全球最重要、最基础的科学难题。

[NirViaje]

• Learning agile and dynamic motor skills for legged robots, ethz, 2019
• 波士顿动力有对手了：不怕摔倒的机器狗，怎么踹都能站起来

[NirViaje]

Deep self modeling, Hod Lipson, 2019

• https://www.creativemachineslab.com/deep-self-modeling.html
• https://engineering.columbia.edu/press-releases/lipson-self-aware-machines
• Task-agnostic self-modeling machines, Hod Lipson, 2019_10.1126@scirobotics.aau9354.pdf

[NirViaje]

该网站将 ArXiv 上的最新机器学习论文与 GitHub 上的代码（TensorFlow/PyTorch/MXNet 等）对应起来。用户可以按标题关键词查询，或者按流行程度、GitHub 收藏数排列「热门研究」。

• https://paperswithcode.com/sota
• AI所有领域最优论文+代码查找神器：966个ML任务、8500+论文任你选

[NirViaje]

人脸检测江湖的那些事儿——从旷视说起

• https://zhuanlan.zhihu.com/p/56619497
  • 尺度变化： SFace
  • 遮挡：FAN
  • 定位精度：2018 WIDER Challenge Face Detection

[NirViaje]

Distilling Free-Form Natural Laws from Experimental Data

• http://science.sciencemag.org/content/324/5923/81
• http://www-conf.slac.stanford.edu/statisticalissues2012/talks/Hod_Lipson.pdf

[NirViaje]

Learning a Generative Model for multi-Step Human-Object Interactions from Videos

• http://www.pirk.info/papers/Wang.etal-2019-LearningInteractions.pdf
• https://mp.weixin.qq.com/s/dyGYlvSivqgVpxpd1XrclQ
• http://www.pirk.info/projects/learning_interactions/index.html
• http://ai.stanford.edu/~hewang/

[NirViaje]

Learning the signatures of the human grasp using a scalable tactile glove

• https://www.nature.com/articles/s41586-019-1234-z
• https://mp.weixin.qq.com/s/Dz3vWVNuSdsLpvgWD1RkHQ
• Learning the signatures of the human grasp using a scalable tactile glove_10.1038@s41586-019-1234-z.pdf
• tbid=563813515898

[NirViaje]

• 最聪明玻璃诞生：以光散射为核心算法，无需耗电，可识别数字 | 专访
• https://www.osa-opn.org/home/newsroom/2019/july/a_different_kind_of_smart_glass/
• Nanophotonic media for artificial neural inference(4).pdf
  • Nanophotonic Media for Artificial Neural Inference, Zongfu Yu, UWMM, 2018_1810.07815.pdf
  • https://photonics.engr.wisc.edu/

[NirViaje]

• Interactive Design Space Exploration and Optimization for CAD Models, Adriana Schulz, MIT, 2017_a11-schulz.pdf

• https://people.csail.mit.edu/aschulz/optCAD/index.html
• https://www.youtube.com/watch?v=45YLK7vbL3M
• https://www.youtube.com/watch?v=_vHMFXfTq6o

• http://cfg.mit.edu/sites/cfg.mit.edu/files/learning_to_fly.pdf

• http://cfg.mit.edu/content/inversecsg-automatic-conversion-3d-models-csg-trees
• https://www.youtube.com/watch?v=mf7Xd6oxNrM
• InverseCSG- Automatic Conversion of 3D Models to CSG Trees, Adriana Schulz, MIT, 2018_inverse_csg.pdf

People

• Computational Design for the Next Manufacturing Revolution, MIT thesis, Adriana Schulz, 2018
  • https://homes.cs.washington.edu/~adriana/
• https://people.csail.mit.edu/wojciech/

[NirViaje]

SDM-NET: Deep Generative Network for Structured Deformable Mesh

• https://arxiv.org/pdf/1908.04520.pdf
• http://geometrylearning.com/sdm-net/

[NirViaje]

DisneyResearch: Vibration-Minimizing Motion Retargeting for Robotic Characters

• https://la.disneyresearch.com/publication/publication-process-vibration-minimizing-motion-retargeting-for-robotic-characters/

  • Vibration-Minimizing Motion Retargeting for Robotic Characters, Disney Research, SIGGRAPH2019_VibSupPaper_2.pdf

• https://mp.weixin.qq.com/s/mh8j_8aqbTGjk_xEVKFuWQ

[NirViaje]

Mirroring without Overimitation

• https://www.jiqizhixin.com/articles/2019-01-30-9
  • https://mp.weixin.qq.com/s/ClowUXgra7CTPcaP64F3lQ 第二篇文章是著名的 UCLA 朱松纯教授和他的合作者贡献的文章「Mirroring without Overimitation: Learning Functionally Equivalent Manipulation Actions」。但这个研究跟朱教授擅长的计算机视觉有联系但不是完全相关，而是利用示教（LfD（learning from demonstration）），令机器人模拟人的镜像神经元（Mirror Neuron）运作进而来控制机器人的对物品的操作能力。当然 LfD 和镜像神经元都不是最创新的概念，但该文跟之前的 LfD 不同之处是提出了机器人的示教学习应该「功能等同性」(functionally equivalent)。具体来说，机器人模拟人的动作，不需要完全学习每个细节动作（overimitation），而只需要学习完成对应的目标。因此在本文的上下文中，「目标」(goal) 被定义为目标对象的期望状态并该状态应该语法模型中编码。
• https://wellyzhang.github.io/blog/2018/12/04/mirroring/
• https://www.seas.upenn.edu/~cffjiang/research/mirror/mirroring2019aaai.pdf
  • http://www.stat.ucla.edu/~sczhu/papers/Conf_2019/AAAI2019_Mirroring_Actions.pdf
  • Mirroring without Overimitation- Learning Functionally Equivalent Manipulation Actions_mirroring2019aaai.pdf

[NirViaje]

• http://taichi.graphics/mpm_course2019/
• https://zhuanlan.zhihu.com/p/97700605
• https://www.youtube.com/watch?v=wKw8LMF3Djo
• https://news.mit.edu/2019/model-design-control-robots-target-1122
• https://www.bilibili.com/video/av81369261/
• http://taichi.graphics/me/
• mls-mpm-cpic.pdf
  • Taichi- A Language for High-Performance Computation on Spatially Sparse Data Structures, YUANMING HU, MIT CSAIL, 2018_taichi_lang.pdf
  • taichi_slides.pdf
  • http://taichi.graphics/mpm_course2019/
• Narrow-Band Topology Optimization on a Sparsely Populated Grid, MIT, UWM, 2018_narrowband_topopt.pdf
• 太極：MIT华人学神开源的计算机图形库
• 《冰雪奇缘》造雪花的技术，被MIT用来开发了一只软体机器人

MIT在读博士生胡渊鸣: Taichi编程语言—高性能稀疏视觉计算与可微编程 网盘（含pdf）>>

• https://pan.baidu.com/s/1wV_0R2V9rMcOyQlw7N6fpw B站>>
• https://www.bilibili.com/video/av82789606/

• [Reference2012] Halide: Decoupling Algorithms from Schedules For High-Performance Image Processing
  • https://cacm.acm.org/magazines/2018/1/223877-halide/fulltext
  • https://github.com/halide/Halide
  • https://halide-lang.org/
  • Halide_CVPR_intro.pdf
  • CVPR_Halide_RecursiveFiltering.pdf

DiffTaichi论文摘要：(ICLR 2020) 基于Taichi，我们提出可微编程语言DiffTaichi，用于构建端到端可微程序。和目前常用的可微编程工具如TensorFlow、PyTorch相比，DiffTaichi更适合构建比常用操作（如卷积、BN等）更不规则的可微运算符，比如可微物理引擎中的粒子网格交互，网格采样等等。DiffTachi的自动微分系统使用“两个尺度”设计：底层通过源代码变换保持并行性和算术强度(arithmetic intensity)，上层通过一个轻量级的磁带(Tape)来记录大内核(Megakernel)的启动。由于省去了枯燥的手动求导过程，DiffTaichi程序比CUDA短4.2倍并具有相同的性能；同时由于其Megakernel的设计，在编写复杂可微程序时，DiffTaichi比TensorFlow快188倍、比PyTorch快13.4倍。我们用DiffTaichi实现了10个不同的可微物理引擎。将神经网络嵌入其中，控制器优化可以在几十个梯度下降迭代中完成，这比增强学习(RL)收敛速度快若干数量级。项目主页和代码：

• https://github.com/yuanming-hu/difftaichi

"This decoupling of data structures from computation makes it easy to experiment with different data structures without changing computation code, and allows users to write computation as if they are working with a dense array, states abstract. "Our compiler then uses the semantics of the data structure and index analysis to automatically optimize for locality, remove redundant operations for coherent accesses, maintain sparsity and memory allocations, and generate efficient parallel and vectorized instructions for CPUs and GPUs."

• A New Language for Sparse Data

[NirViaje]

Michael Levin, Josh Bongard

• Automated shapeshifting for function recovery in damaged robots, Levin, Josh Bongard, 2019_p28.pdf
• A scalable pipeline for designing reconfigurable organisms, Levin, Josh Bongard, 2020, 10.1073@pnas.1910837117.pdf
• https://github.com/skriegman/reconfigurable_organisms

[NirViaje]

解决奖励预测问题的一个重要突破是时间差学习（TD）算法，该算法不会去计算未来的总回报，而只是尝试在下一个时刻预测即时奖励。

当下一刻出现新信息时，将新的预测与预期的进行比较。如果二者不同，则计算出回报的差异，并使用此“即时差异”将旧预测调整为新预测，使期望预测与现实相匹配，整个预测链逐渐变得更加准确。

• DeepMind哈佛惊人发现！大脑中竟存在“分布强化学习”，靠多巴胺驱动
  • https://deepmind.com/blog/article/Dopamine-and-temporal-difference-learning-A-fruitful-relationship-between-neuroscience-and-AI
  • https://www.nature.com/articles/s41586-019-1924-6
  • https://osf.io/ux5rg/
  • A distributional code for value in dopaminebased reinforcement learning, 2019, DeepMind_10.1038@s41586-019-1924-6.pdf

[NirViaje]

SLIDE : IN DEFENSE OF SMART ALGORITHMS OVER HARDWARE ACCELERATION FOR LARGE-SCALE DEEP LEARNING SYSTEMS

• https://www.cs.rice.edu/~as143/Papers/SLIDE_MLSys.pdf
• https://github.com/keroro824/HashingDeepLearning
• 学习超大神经网络，CPU超越V100 GPU，靠的居然是哈希？
  • https://mp.weixin.qq.com/s/VaJVEufFWdVm-Lb32sIRPw 如此神奇的加速是怎么实现的？具体而言，研究者采用了局部敏感哈希（Locality Sensitive Hashing）算法，并在神经网络中使用了自适应 dropout。局部敏感哈希是一类哈希算法，当输入数据彼此类似的时候，具有更高的碰撞概率，而不相似的算法彼此碰撞的概率很低。一种广泛应用的最近邻逼近搜索算法就使用了局部敏感哈希理论。

[NirViaje]

Memristive Synapses for Brain-Inspired Computing, THU, 2019

• 清华团队研制出全球首款多阵列忆阻器存算一体系统
  • https://mp.weixin.qq.com/s/1yGrDKS-sr_EsGylv9k0Qw
  • [Uploading Memristive Synapses for Brain-Inspired Computing_wang2019.pdf…]()

[NirViaje]

• https://arxiv.org/pdf/2012.03806.pdf
• Perspectives on Sim2Real Transfer for Robotics- A Summary of the R-SS 2020 Workshop_2012.03806.pdf

Model-based RL

Introducing Dreamer: Scalable Reinforcement Learning Using World Models

• World Models -> PlaNet -> Dreamer
• https://ai.googleblog.com/2020/03/introducing-dreamer-scalable.html
• https://danijar.com/project/dreamer/
• https://medium.com/dataseries/google-deepminds-dreamer-is-a-reinforcement-learning-agent-that-can-solve-long-horizon-tasks-5faa6f6b63b
• https://arxiv.org/pdf/1912.01603.pdf
  • DREAM TO CONTROL- LEARNING BEHAVIORS BY LATENT IMAGINATION, deepmind, 2020_1912.01603.pdf
• https://danijar.com/material/2020-dreamer-slides.pdf
  • Dream to Control_2020-dreamer-slides.pdf

Learning Latent Dynamics for Planning from Pixels

• https://arxiv.org/pdf/1811.04551.pdf

Self-Modeling

• https://www.creativemachineslab.com/deep-self-modeling.html
• Task-agnostic self-modeling machines, Hod Lipson, 2019_scirobotics.aau9354.pdf

Truss

Hod Lipson works with truss

• https://www.creativemachineslab.com/truss-reconfiguration.html

Learn from rendering

• https://jiajunwu.com/papers/dissertation.pdf
• https://jiajunwu.com/

; https://www.cs.columbia.edu/~shurans/ ; http://robotouch.ri.cmu.edu/yuanwz/gelsight-projects.htm ; https://viterbi-web.usc.edu/~limjj/ ; https://www.clvrai.com/publications/ ; http://www.cs.columbia.edu/~bchen/, with Hod Lipson ; Ikea