三种分布式人工智能主流框架方案对比

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在最近的几年中，人工智能随着神经网络的突破，得到了巨大的发展，特别在图像、分析、推荐等领域。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在人工智能快速发展的同时，计算规模不断扩大、专家系统过於单一、神经网络模型的灵活性、应用领域的复杂行等问题，也在不断升级。在这样的环境下，分布式人工智能的发展被研究机构和大型企业提上日程。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"分布式人工智能，可以解决集中化人工智能的几个主要问题："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"规模化的计算问题。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"计算模型的拆分训练。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"多智能专家系统的协作。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"多智能体博弈和训练演化，解决数据集不足的问题。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"群体智能决策和智能系统决策树的组织，适应复杂的应用场景，比如工业、生物、航天等领域。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"适应物联网和小型智能设备，联合更多的计算设备和单元。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"针对现有的分布式人工智能组件和框架的研究，可以看下几个主流框架的方案。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"  "}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"Tensorflow 的分布式方案"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Tensorflow 底层数据是高性能的 gRPC 库，分布式集群的组件主要是三个部分：client、master、worker process。运用它的组件，可以形成两种主要的分布式部署模式："},{"type":"text","marks":[{"type":"strong"}],"text":"单机多卡（single process）"},{"type":"text","text":"和"},{"type":"text","marks":[{"type":"strong"}],"text":"多机多卡（multi-device process）"},{"type":"text","text":"。方案出于 Martin Abadi、Ashish Agarwal、Paul Barham 论文《TensorFlow:Large-Scale Machine Learning on Heterogeneous Distributed Systems》。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"它的部署和运作模式，如图 1-1 所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/6e\/6eea4f9c75a156b01bce241236d33285.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图 1-1 Tensorflow 单机单卡与多级多卡部署模式"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":"br"}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"对于普通的机器学习，单机单卡就可以进行运算，在运算规模不断升级的时候，就需要"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"对 tensoflow 的训练设计分布式结构。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"机器学习在的参数训练，主要是两大过程："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用卷积参数训练梯度"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"根据梯度再优化更新参数"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在大规模的计算过程中，就需要进行集群计算，在小型规模的时候，可以使用单个机器，多个 CPU\/GPU 来进行计算，在规模越来越大的时候，可以使用多个机器进行并行计算。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在 tensorflow 中分布式集群定义为一个“集群任务”（tf.train.ClusterSpec），将任务分布在不同的“服务器”（tf.train.Server）中，服务器中包含了“master”运行进程和”worker“计算进程。在分布式集群运算中，主要有以下结构："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Task"},{"type":"text","text":"：分布式中的计算任务，每个 job 结构中，每个任务拥有唯一索引。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Job"},{"type":"text","text":"：每个 Job 表示了一个完整的训练目标，其中的 Task 都有唯一的目标，为同一个训练目标进行任务拆分。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Cluser"},{"type":"text","text":"：集群集合，其中管理了多个 Job，一般一个集群对应了一个专门的神经网络，其中不同的 job 负责了不同的目标，比如梯度计算、参数优化、任务训练等。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Tensorflow Server"},{"type":"text","text":"：处理服务节点过程，并且对外提供master server 和  worker server 的接口服务。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Master Server"},{"type":"text","text":"：和远程分布式设备进行交互的服务，协调多个 worker server 工作。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Worker Service"},{"type":"text","text":"：Tensorflow 图计算过程中，任务进程处理的服务，单独运行各自的任务。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Client"},{"type":"text","text":"：客户端，用来管理和启动 AI 工作，与远程服务集群交互。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Tensorflow 的多机多卡分布式模式，主要有以下几种："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"In-graph 模式"},{"type":"text","text":"：在多机多卡的模式下，各自的节点和运算服务，都通过单一的节点进行数据分发，拥有集中化的管理方式，最终数据归并到统一的节点上。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Between-graph 模式"},{"type":"text","text":"：训练的数据保存在各自的节点上，不进行分发。每个节点的权重都是平等的，当计算节点各自计算完以后，需要告诉参数服务器，并且更新优化参数，是比较推荐的分布式方式。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"根据 Tensorflow 的分布式机器学习方案，我们可以总结它存在的一些缺点，如下所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Tensorflow 在分布式的计算方式中，分为了训练和参数更新服务，极大得扩展了规模化的计算能力，但是在数据归并和多级分层上，并没有提供非常好的解决方案。没有较好的计算模型，去提供多层次的分布式结构。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Tensorflow 在分布式的计算模型中，主要分为了训练任务、梯度计算、参数更新等，但是确没有提供一个较好的计算模型构造方式，去提供分布式 AI 所需要的任务分发、数据归并、模型更新等能力。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Tensorflow 在分布式的组织方式上，更像是一个平级层次的分布式，最终统一归并成为一个最终模型，它最终完成的是一个集中化的单一专家系统，无法提供群体智能能力，也无法形成多群体博弈对抗的协调决策，无法支持多专家系统的兼容和协调。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Tensorflow 需要运行在高性能计算机，高性能的 GPU 下，很难支持小型设备、物联网等多设备的边缘化集群计算。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"Spark 分布式机器学习"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上面提到，Tensorflow 在机器学习并行计算的多层结构上，没有较好的解决方案，Spark 根据它在大数据领域的分布式框架，很好的解决了分层任务划分的处理。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在 Spark 中，计算模型可以设计成为有向无环图 DAG，无环图的顶点是 RDD，它是 Spark的核心组件。RDD 是一种弹性的分布式数据集，它可以支持多个 RDD 分片的依赖、变换（transformation）和动作 (action)，它可以从从 RDD A 变换成为 RDD B，这个 transformation 就是 DAG 无环图的边。通过 DAG 的表示方法，表示计算模型，并且编译成为 Stage。这个模型的表示，如图 1-2 所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/15\/154baf0780bb0d456271250e8b4a2631.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图 1-2 Spark 计算模型和 Stage 形成的 DAG 图"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Spark 运用它的分布式机制，可以支持多个 Stage 并行计算，也能支持 Stage 下属 Stage 模型分层结构。Spark 基于它的 master-worker 架构，可以将 DAG 中的 stage 分割，指定到不同的机器上面执行任务。它的驱动器 driver 负责协调任务和调度器组件 scheduler，调度器分为 DAG 和 Task 调度器，用来分配 Task 到不同的运算单元。它的分布式并行运算流程，如图 1-3 所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/0c\/0c014f3ac2b76983836863a22c611556.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图 1-3 Spark 分布式并行计算分层"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"上图表示了在分布式 AI 的训练环节，通过 Driver 规划整体模型，通过节点 Executor 端分别计算，再回归汇总给 Driver，它们的具体要点如下所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"一个具体的 AI 系统总结为一个工作流 Work Process，其中包含了许多 action 算子操作。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"action 算子将不同的工作类别划分为不同 Job。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Job 根据模型的宽以来，划分给不同的 Stage 小模型。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Stage 的内部也可以再划分成为多个小型的 Stage 模型，如图 1-2 所示。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Stage 将功能相同的，DAG 中的边运算归纳为不同的 Task。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Task 可以提交 Executor 做最小运算单元的运算，最终反馈给 Driver 进行模型更新和存储。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"通过 Spark 的分布式计算框架与机器学习 caffe 运算库进行结合，伯克利大学 Michael I. Jordan 组发布了论文《SparkNet: Training Deep Network in Spark》，他们开发了 SparkNet 库，它是基于 Spark 的深度神经网络架构，主要能力如下所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"提供易用的，用于神经网络的接口，可以快速访问 RDD。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"提供 Scala 的接口，可以与 caffe 进行交互。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"提供轻量级的卷积 tensor 库。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"提供简单的并行机制和通讯机制。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"即插即拔，易于部署。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"兼容现有的 caffe 模型。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"根据 Spark 的分布式机器学习机制，我们可以总结它存在的一些缺点，如下所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Spark 支持大规模的模型计算分层，但是计算消耗较大，不适合用在较小规模的分布式方案。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Spark 在节点和节点之间的数据路由不是非常灵活。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Spark 难以在小型化设备进行联合学习运算，无法部署在小型化设备之上。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Spark 的框架是一个并行计算框架，它的模式是任务确认调度-模型划分-分片计算-归并总结，更适合大数据集的机器学习，不适合多节点博弈等强化学习方法。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"Google 联合学习方案"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"针对小型化设备，google 在一篇文章《Federated Learning: Collaborative Machine"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Learning without Centralized Training Data》中，提出了联邦联合学习的概念。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"它的工作原理如下所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":null,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"手机或者其他小型设备，下载云端的共享模型。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"每个小型设备的用户，通过自己的历史数据来训练和更新模型。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"将用户个性化更新后的模型，抽取成为一个小的更新文件。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":4,"align":null,"origin":null},"content":[{"type":"text","text":"提取模型的差异化部分，进行加密，上传到云端。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":5,"align":null,"origin":null},"content":[{"type":"text","text":"在云端将新用户的差异化模型和其他用户模型，进行 Average 平均化，然后更新改善现有共享模型。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这样工作的好处有如下部分："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"聚合了边缘的小型化设备（比如手机），增加了 AI 的数据来源和计算能力来源。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在机器学习的模型结果上，更加适应广泛用户的行为数据模型。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"不同群体之间，可以产生博弈和不断强化的模型结果，模型可以在广泛的分布式基础下，不断迭代更新。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本文转载自：金科优源汇（ID：jkyyh2020）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文链接："},{"type":"link","attrs":{"href":"https:\/\/mp.weixin.qq.com\/s\/t7zS7RwjW4j3m2SOllrL1A","title":"xxx","type":null},"content":[{"type":"text","text":"三种分布式人工智能主流框架方案对比"}]}]}]}