[{"data":1,"prerenderedAt":489},["ShallowReactive",2],{"content-query-b3dZDUVYov":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"state":13,"body":14,"_type":483,"_id":484,"_source":485,"_file":486,"_stem":487,"_extension":488},"/activities/zh/254","zh",false,"","MindSpore社区8月新社区群组介绍系列之四","MindSpore着眼面向未来共同技术研究AI课题，面向十大方向及其简介","2020-09-10","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/09/21/32723a8f9b484ed28f378b3bd5012ed8.png","activities",1,{"type":15,"children":16,"toc":473},"root",[17,25,32,35,41,46,51,56,64,69,74,79,84,89,94,99,104,109,114,119,124,129,134,139,144,149,154,159,164,169,174,179,184,189,194,199,204,209,214,219,224,229,234,239,244,249,253,258,262,267,271,276,280,285,289,294,298,303,307,312,316,321,325,330,335,340,344,349,354,359,366,380,385,396,401,412,417,428,439,444,454,461,466],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"mindspore社区8月新社区群组介绍系列之四",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":29},"h2",{"id":28},"wg-research",[30],{"type":24,"value":31},"WG-Research",{"type":18,"tag":26,"props":33,"children":34},{"id":7},[],{"type":18,"tag":36,"props":37,"children":38},"p",{},[39],{"type":24,"value":40},"人工智能领域",{"type":18,"tag":36,"props":42,"children":43},{},[44],{"type":24,"value":45},"作为一个技术发展日新月异，新研究新成果层出不穷的领域，MindSpore社区持续关注着人工智能领域发展的新方向，并总结出10大影响人工智能发展的方向，针对这10大方向成立WG-Research（研究工作组），在了解了WG-Research之后，希望有幸能邀请您一起进行深度探讨~",{"type":18,"tag":36,"props":47,"children":48},{},[49],{"type":24,"value":50},"简单介绍完毕让我们进一步了解MindSpore社区系列之四——WG-Research",{"type":18,"tag":36,"props":52,"children":53},{},[54],{"type":24,"value":55},"所面向的是哪10大方向，针对这10个课题又有什么idea呢？",{"type":18,"tag":36,"props":57,"children":58},{},[59],{"type":18,"tag":60,"props":61,"children":63},"img",{"alt":7,"src":62},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/09/10/a16aada3bb5c41e5b3c30bb003c83df2.gif",[],{"type":18,"tag":36,"props":65,"children":66},{},[67],{"type":24,"value":68},"MindSpore着眼面向未来共同技术研究AI课题",{"type":18,"tag":36,"props":70,"children":71},{},[72],{"type":24,"value":73},"NO.1 通用的张量可微分计算框架",{"type":18,"tag":36,"props":75,"children":76},{},[77],{"type":24,"value":78},"· 新网络模型对深度学习框架的IR表达、优化和执行提出挑战，包括引入新的op抽象级别，模型的动态性等;",{"type":18,"tag":36,"props":80,"children":81},{},[82],{"type":24,"value":83},"· 第三方高性能计算语言或框架加速，迫切需要通用性更强的张量计算框架和API设计;",{"type":18,"tag":36,"props":85,"children":86},{},[87],{"type":24,"value":88},"· 模型层和算子层统一优化的技术挑战，包括层次化IR设计、优化基础设施、自动化调优、循环优化等。",{"type":18,"tag":36,"props":90,"children":91},{},[92],{"type":24,"value":93},"NO.2 分布式并行原生的AI计算框架",{"type":18,"tag":36,"props":95,"children":96},{},[97],{"type":24,"value":98},"· 模型规模和复杂度越来越高。比如1750亿参数量的GPT-3 ，百亿特征推荐等；",{"type":18,"tag":36,"props":100,"children":101},{},[102],{"type":24,"value":103},"· 模型手工切分难。需要结合计算量、集群规模、通信带宽、网络拓扑等信息才能构造出并行模式；",{"type":18,"tag":36,"props":105,"children":106},{},[107],{"type":24,"value":108},"· 并行表达缺乏适应性。单纯Graph级别的模型切分无法获得有效加速比，需要算法逻辑和并行逻辑解耦。",{"type":18,"tag":36,"props":110,"children":111},{},[112],{"type":24,"value":113},"NO.3 整网低bit训练策略",{"type":18,"tag":36,"props":115,"children":116},{},[117],{"type":24,"value":118},"目前混合精度能自动针对网络进行fp16、fp32进行精度调节以提升训练性能和内存优化，但这种优化针对不同的ai芯片策会有所不同（算子开销情况不同），不同硬件不同网络的配置有所差异。",{"type":18,"tag":36,"props":120,"children":121},{},[122],{"type":24,"value":123},"因此，如何自动的生成适配各种硬件的精度调节策略，特别是更激进的低bit策略和更大范围的算子降精度。",{"type":18,"tag":36,"props":125,"children":126},{},[127],{"type":24,"value":128},"NO.4 内存优化",{"type":18,"tag":36,"props":130,"children":131},{},[132],{"type":24,"value":133},"内存有多种优化策略，有些优化策略基于重计算、host-device内存倒换等，会增加计算量。但能通过增加计算量的进一步突破内存瓶颈，提高网络的batchsize，而batchsize的提升往往能提升GPU、NPU的利用率从而提升吞吐性能。",{"type":18,"tag":36,"props":135,"children":136},{},[137],{"type":24,"value":138},"NO.5 可验证可信AI",{"type":18,"tag":36,"props":140,"children":141},{},[142],{"type":24,"value":143},"· 可信AI（或AI责任）的多个方面，例如鲁棒性、后门、公平性、隐私保护能力、可追责性等，逐渐引起业界关注；",{"type":18,"tag":36,"props":145,"children":146},{},[147],{"type":24,"value":148},"· 学者对于可信AI属性的理解和研究是许多是基于实证的（empirical），基于理论的研究较少；",{"type":18,"tag":36,"props":150,"children":151},{},[152],{"type":24,"value":153},"· 可信AI属性约束的可验证、可证明的分析、调优、评价方法需要理论指导。",{"type":18,"tag":36,"props":155,"children":156},{},[157],{"type":24,"value":158},"NO.6 密态AI",{"type":18,"tag":36,"props":160,"children":161},{},[162],{"type":24,"value":163},"· AI业务在训练和部署过程中，数据、模型和计算资源可能分属各方，因此涉及大量的数据跨信任域移动，数据隐私保护和模型机密性保护的问题突出。",{"type":18,"tag":36,"props":165,"children":166},{},[167],{"type":24,"value":168},"· 密态计算是保护关键数据机密性的重要方向。目前，基于可信执行环境的密态计算具有性能优势，然而信任模型受限；基于加密学（如同态加密，多方计算）的密态计算信任模型简单，然而性能距离实用尚存在差距；",{"type":18,"tag":36,"props":170,"children":171},{},[172],{"type":24,"value":173},"· AI场景下的一系列针对性优化可能提升密态AI的性能，包括但不限于：适应AI场景的密态算法和神经网络结构、中间层表达、编译优化技术、基于硬件的优化加速。",{"type":18,"tag":36,"props":175,"children":176},{},[177],{"type":24,"value":178},"NO.7 自动调优推荐",{"type":18,"tag":36,"props":180,"children":181},{},[182],{"type":24,"value":183},"训练出满足精度要求的模型，往往需要丰富的专家知识以及反复迭代尝试。虽然有AutoML技术，但仍然存在搜索空间设定难、搜索空间大训练时间长的问题。如果可以结合用户训练的迭代历史，通过分析历史数据，实现一种轻量化的超参推荐手段。",{"type":18,"tag":36,"props":185,"children":186},{},[187],{"type":24,"value":188},"同样，对于性能调优，也存在类似问题，在不同的异构硬件、模型、数据处理场景下，需要专家知识进行调优。因此是否可以通过自动识别系统性能瓶颈、推荐代码路径等方法，来降低性能调优门槛？",{"type":18,"tag":36,"props":190,"children":191},{},[192],{"type":24,"value":193},"NO.8 符合人类认知的可解释AI方法",{"type":18,"tag":36,"props":195,"children":196},{},[197],{"type":24,"value":198},"目前的深度学习模型因为其技术原理存在本质上的黑盒性质，从而导致AI服务的不透明及不可解释性，进一步约束了基于AI服务的商业应用与推广。现有的可解释AI技术主要关注于如何对黑盒模型进行有限度的、工程性的辅助信息输出技术，而忽略了从人类认知角度，如何对AI模型进行理解。",{"type":18,"tag":36,"props":200,"children":201},{},[202],{"type":24,"value":203},"人类理解事物通常通过类比、比喻、归纳等认知方法，并且具有一定的心理认知构建流程。期望能够探寻比较体系化的符合人类认知的可解释AI方法，包括交互界面、解释方法、量度手段等等。",{"type":18,"tag":36,"props":205,"children":206},{},[207],{"type":24,"value":208},"NO.9 NN与机器学习结合的通用AI",{"type":18,"tag":36,"props":210,"children":211},{},[212],{"type":24,"value":213},"结合传统模型和神经网络的模型创新是研究热点",{"type":18,"tag":36,"props":215,"children":216},{},[217],{"type":24,"value":218},"深度概率模型创新：通过神经网络和概率模型的组合，该模型可以更好的表达不确定性地帮助决策",{"type":18,"tag":36,"props":220,"children":221},{},[222],{"type":24,"value":223},"图神经网络：神经网络与传统的图结构相结合，面向认知推理和未来趋势",{"type":18,"tag":36,"props":225,"children":226},{},[227],{"type":24,"value":228},"NO.10 AI和科学计算聚合创新",{"type":18,"tag":36,"props":230,"children":231},{},[232],{"type":24,"value":233},"AI建模：AI自动建模可以有效地提高科学计算的建模效率，收敛性分析可以提高模型的可靠性，并确保用户的简单放心使用。",{"type":18,"tag":36,"props":235,"children":236},{},[237],{"type":24,"value":238},"AI解决方案：高阶微分的计算量随参数和阶数呈指数增长。我们可以设计神经网络模型来解决此类经典问题。",{"type":18,"tag":36,"props":240,"children":241},{},[242],{"type":24,"value":243},"发起WG-Research（研究工作组）",{"type":18,"tag":36,"props":245,"children":246},{},[247],{"type":24,"value":248},"华为的MindSpore社区团队决定在社区发起研究工作组，面向以上十大方向，征集大家的智慧与能力，开展创新协作：",{"type":18,"tag":36,"props":250,"children":251},{},[252],{"type":24,"value":73},{"type":18,"tag":36,"props":254,"children":255},{},[256],{"type":24,"value":257},"前沿应用驱动，从新型模型、动态模型、高性能计算等视角，研究未来计算框架的演进方向和关键技术路径。",{"type":18,"tag":36,"props":259,"children":260},{},[261],{"type":24,"value":93},{"type":18,"tag":36,"props":263,"children":264},{},[265],{"type":24,"value":266},"以超大模型为驱动，研究加速分布式训练的关键技术，包括但不局限于自动并行、混合并行、内存优化、弹性伸缩等。",{"type":18,"tag":36,"props":268,"children":269},{},[270],{"type":24,"value":113},{"type":18,"tag":36,"props":272,"children":273},{},[274],{"type":24,"value":275},"自适应的提供针对各种网络，整网低bit精度训练机制。",{"type":18,"tag":36,"props":277,"children":278},{},[279],{"type":24,"value":128},{"type":18,"tag":36,"props":281,"children":282},{},[283],{"type":24,"value":284},"如何自适应的找到合适的内存优化策略使得在网络能执行的前提下，整体性能优化或者提供一种策略。",{"type":18,"tag":36,"props":286,"children":287},{},[288],{"type":24,"value":138},{"type":18,"tag":36,"props":290,"children":291},{},[292],{"type":24,"value":293},"对于可信AI，提出可验证、可证明的机制研究、评价体系等。",{"type":18,"tag":36,"props":295,"children":296},{},[297],{"type":24,"value":158},{"type":18,"tag":36,"props":299,"children":300},{},[301],{"type":24,"value":302},"达成一个实际AI应用场景下性能可行、灵活的AI密态计算框架，或者其中的关键技术。",{"type":18,"tag":36,"props":304,"children":305},{},[306],{"type":24,"value":178},{"type":18,"tag":36,"props":308,"children":309},{},[310],{"type":24,"value":311},"可以自动推荐优化的超参配置、性能优化路径，降低模型开发、使用门槛，提升模型调试调优效率。",{"type":18,"tag":36,"props":313,"children":314},{},[315],{"type":24,"value":193},{"type":18,"tag":36,"props":317,"children":318},{},[319],{"type":24,"value":320},"一套完整的符合人类认知的可解释AI方法策略，针对不同场景、不同认知，提供需要的交互认知接口设计方案，并对典型场景的有个例设计。",{"type":18,"tag":36,"props":322,"children":323},{},[324],{"type":24,"value":208},{"type":18,"tag":36,"props":326,"children":327},{},[328],{"type":24,"value":329},"完备的概率采样库和概率推断（通过已知样本学习总体样本的概率分布）算法库；",{"type":18,"tag":36,"props":331,"children":332},{},[333],{"type":24,"value":334},"面向动态变化的异构图（特征维度不同，信息汇聚方式不同）设计新的算法；",{"type":18,"tag":36,"props":336,"children":337},{},[338],{"type":24,"value":339},"超大规模（Trillion）分布式图数据存储、切分和采样。",{"type":18,"tag":36,"props":341,"children":342},{},[343],{"type":24,"value":228},{"type":18,"tag":36,"props":345,"children":346},{},[347],{"type":24,"value":348},"建模：针对科学问题构造神经网络，训练数据和Loss函数",{"type":18,"tag":36,"props":350,"children":351},{},[352],{"type":24,"value":353},"求解：AI模型解微分方程，解优化问题，支持高阶微分",{"type":18,"tag":36,"props":355,"children":356},{},[357],{"type":24,"value":358},"如何参与WG-Research",{"type":18,"tag":360,"props":361,"children":363},"h3",{"id":362},"_1-熟悉课题计划",[364],{"type":24,"value":365},"1. 熟悉课题计划",{"type":18,"tag":367,"props":368,"children":369},"ul",{},[370],{"type":18,"tag":371,"props":372,"children":373},"li",{},[374],{"type":18,"tag":375,"props":376,"children":377},"strong",{},[378],{"type":24,"value":379},"2020-9",{"type":18,"tag":36,"props":381,"children":382},{},[383],{"type":24,"value":384},"MindSpore TSC评议research working plan",{"type":18,"tag":367,"props":386,"children":387},{},[388],{"type":18,"tag":371,"props":389,"children":390},{},[391],{"type":18,"tag":375,"props":392,"children":393},{},[394],{"type":24,"value":395},"2020-10",{"type":18,"tag":36,"props":397,"children":398},{},[399],{"type":24,"value":400},"创建research working group并公开发布课题内容",{"type":18,"tag":367,"props":402,"children":403},{},[404],{"type":18,"tag":371,"props":405,"children":406},{},[407],{"type":18,"tag":375,"props":408,"children":409},{},[410],{"type":24,"value":411},"2020-10~2021-10",{"type":18,"tag":36,"props":413,"children":414},{},[415],{"type":24,"value":416},"2. 查看WG-Research的立项文档：",{"type":18,"tag":36,"props":418,"children":419},{},[420],{"type":18,"tag":421,"props":422,"children":426},"a",{"href":423,"rel":424},"https://gitee.com/mindspore/community/blob/master/design/meps/mep-research/MEP-RESEARCH.md",[425],"nofollow",[427],{"type":24,"value":423},{"type":18,"tag":36,"props":429,"children":430},{},[431,433],{"type":24,"value":432},"如果你有任何问题，欢迎邮件联系工作组负责人",{"type":18,"tag":421,"props":434,"children":436},{"href":435},"mailto:wang1@huawei.com",[437],{"type":24,"value":438},"wang1@huawei.com",{"type":18,"tag":36,"props":440,"children":441},{},[442],{"type":24,"value":443},"3. 在社区提交ISSUE或者PR，正式开始提出你的问题或者做出你的贡献",{"type":18,"tag":36,"props":445,"children":446},{},[447,449],{"type":24,"value":448},"4.微信添加好友搜索：",{"type":18,"tag":375,"props":450,"children":451},{},[452],{"type":24,"value":453},"mindspore0328",{"type":18,"tag":36,"props":455,"children":456},{},[457],{"type":18,"tag":60,"props":458,"children":460},{"alt":7,"src":459},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/09/10/a96ec9592c2c46cf96e4763b2f26d952.jpg",[],{"type":18,"tag":26,"props":462,"children":464},{"id":463},"扫描下方二维码关注公众号",[465],{"type":24,"value":463},{"type":18,"tag":36,"props":467,"children":468},{},[469],{"type":18,"tag":60,"props":470,"children":472},{"alt":7,"src":471},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/09/10/619411af53174067b1e3501718fa168f.jpg",[],{"title":7,"searchDepth":474,"depth":474,"links":475},4,[476,478,482],{"id":28,"depth":477,"text":31},2,{"id":7,"depth":477,"text":7,"children":479},[480],{"id":362,"depth":481,"text":365},3,{"id":463,"depth":477,"text":463},"markdown","content:activities:zh:254.md","content","activities/zh/254.md","activities/zh/254","md",1776506019200]