深度学习的未来 #1941

changkun · 2017-07-22T23:13:05Z

Issue: 深度学习的未来 #1930
@sqrthree 初稿完成

sunshine940326 · 2017-07-23T08:33:59Z

校对认领

linhe0x0 · 2017-07-23T08:34:02Z

@sunshine940326 好的呢 🍺

MountainOne · 2017-07-23T11:00:53Z

校对认领

linhe0x0 · 2017-07-23T11:00:55Z

@MoutainOne 妥妥哒 🍻

MountainOne · 2017-07-23T11:11:52Z

TODO/the-future-of-deep-learning.md

-It is part of a series of two posts on the current limitations of deep learning, and its future.
-You can read the first part here: [The Limitations of Deep Learning](https://blog.keras.io/the-limitations-of-deep-learning.html).
+这篇文章改编自我的书 [Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras&a_bid=76564dff)（Manning 出版社）第 9 章第 3 节（译者注：「结论」一章最后一小节）。
+它是目前对深度学习的局限性及其未来的两篇系列文章的第二篇。


它是目前对深度学习的局限性及其未来的两篇系列文章的第二篇。 => 它是《当前深度学习的局限性及其未来》系列文章的第二篇。

MountainOne · 2017-07-23T11:31:29Z

TODO/the-future-of-deep-learning.md

- New forms of learning that make the above possible—allowing models to move away from just differentiable transforms.
- Models that require less involvement from human engineers—it shouldn't be your job to tune knobs endlessly.
- Greater, systematic reuse of previously learned features and architectures; meta-learning systems based on reusable and modular program subroutines.
+- 在比目前可比较的层次更丰富的原始数据之上，构建更贴近通用计算机程序的模型 —— 也就是说我们如何能够得到（数据的）**推理 (reasoning)**和**抽象(abstraction)**，这也是目前模型的根本弱点。


在比目前可比较的层次更丰富的原始数据之上，构建更贴近通用计算机程序的模型 —— 也就是说我们如何能够得到（数据的）推理 (reasoning)和抽象(abstraction)，这也是目前模型的根本弱点。=>
构建于顶层丰富的原始数据之上而不是当前的可微层次的模型，将更加接近通用目的的计算机程序 —— 当前模型的根本弱点是我们将如何得到（数据的）推理 (reasoning)和抽象(abstraction)。

第一遍读下来感觉有点绕口，这样改了一下会不会稍微好些。破折号后面的那句话感觉应该是一个倒装句，正常翻下来完全读不通。

倒桩的译法削弱了原作者对弱点的强调吧？我的理解是破折号后面的内容是对前面「blablabla...构建blablabla的模型」的补充说明，「——也就是说blablabla」，最后一句再进一步补充说明这个blablabla「这也是目前模型的根本弱点」

对，破折号后面的内容是对前面的补充说明，但也是一个完整的陈述句。你可以用“the fundamental weakness of current models”替换掉前面的主语this,然后整个句子就是一个完整的陈述句了，这里之所以用倒装是为了避免整个句子头重脚轻。

个人比较认同 MoutainOne 的想法

MountainOne · 2017-07-23T11:39:44Z

TODO/the-future-of-deep-learning.md

- Models that require less involvement from human engineers—it shouldn't be your job to tune knobs endlessly.
- Greater, systematic reuse of previously learned features and architectures; meta-learning systems based on reusable and modular program subroutines.
+- 在比目前可比较的层次更丰富的原始数据之上，构建更贴近通用计算机程序的模型 —— 也就是说我们如何能够得到（数据的）**推理 (reasoning)**和**抽象(abstraction)**，这也是目前模型的根本弱点。
+- 能够学习新的形式，使上述成为可能 —— 允许模型脱离（每一步之间的）微分变换 (differentiable transformation) 的限制。（译者注：神经网络的每一步传播本质上是一个微分变换）


能够学习新的形式，使上述成为可能 —— 允许模型脱离（每一步之间的）微分变换 (differentiable transformation) 的限制。 => 允许模型摆脱积分变换的新的学习模式使得实现上述的模型成为可能。

这句话也是要翻译成一个倒装句才能读通。

还是不太赞同这个倒装？我觉得作者是先给出结论是未来的深度学习「能够学习新的形式」作为第一点的补充；破折号后面的内容则是进一步解释如何达到这个目标。好像没什么读不通的..

是的，作者的意思应该是这样没错。但感觉不符合中国人的造句习惯，读起来不是很舒服（我是这么觉得），破折号那里处理一下我感觉不那么突兀。

允许模型摆脱积分变换的新的学习模式使得实现上述的模型成为可能。这样翻译读起来更加顺

MountainOne · 2017-07-23T11:44:56Z

TODO/the-future-of-deep-learning.md

- Greater, systematic reuse of previously learned features and architectures; meta-learning systems based on reusable and modular program subroutines.
+- 在比目前可比较的层次更丰富的原始数据之上，构建更贴近通用计算机程序的模型 —— 也就是说我们如何能够得到（数据的）**推理 (reasoning)**和**抽象(abstraction)**，这也是目前模型的根本弱点。
+- 能够学习新的形式，使上述成为可能 —— 允许模型脱离（每一步之间的）微分变换 (differentiable transformation) 的限制。（译者注：神经网络的每一步传播本质上是一个微分变换）
+- 不需要人工调参的模型 —— 你的工作不应该是无休止地测试不同的参数。


不需要人工调参的模型 => 不需要过多人工参与调参的模型

应该是需要更少的人工参与调参，而不是完全不需要人工调参

MountainOne · 2017-07-23T12:56:53Z

TODO/the-future-of-deep-learning.md


-I would definitely expect this subfield to see a wave of renewed interest in the next few years. In particular, I would expect the emergence of a crossover subfield in-between deep learning and program synthesis, where we would not quite be generating programs in a general-purpose language, but rather, where we would be generating neural networks (geometric data processing flows) augmented with a rich set of algorithmic primitives, such as for loops—and many others. This should be far more tractable and useful than directly generating source code, and it would dramatically expand the scope of problems that can be solved with machine learning—the space of programs that we can generate automatically given appropriate training data. A blend of symbolic AI and geometric AI. Contemporary RNNs can be seen as a prehistoric ancestor to such hybrid algorithmic-geometric models.
+我相当期待这个子领域在未来的几年里掀起一股新浪潮。特别地，我期望深度学习和程序合成之间能够再出现一个交叉子领域，在这里我们再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强，比如 for 循环等等。这会比直接生成源代码要容易得多，而且他会大大的扩展机器学习可以解决问题的范围 —— 我们可以自动生成给定适当训练数据的程序空间。一个符号 AI 与几何 AI 的混合。当代的 RNN 可以看做是这种混合算法与几何模型的鼻祖。


在这里我们再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强 =>
在这里我们不再用通用语言来写程序，而是由经过丰富算法原语集增强的神经网络（几何数据处理流）来自动生成

前半句少加了一个'不',后半句“加强”应该是一个形容词而不是动词，用来修饰神经网络，后半句的动作应该是生成程序。（感觉“geneating”后少了一个"through"）

在这里我们再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强 =>
在这里我们不再用通用语言来写程序，而是生成通过丰富的算法原语集增强的神经网络（几何数据处理流）。

又看了下下文，感觉这句话的意思是，生成的是网络，而不是直接生成程序源代码。

有异议，原文上下文已经足够清楚的表明「模型即程序」这个未来愿景，所以此处原作希望表达的意愿应该就是通过神经网来生成程序；此外，argumented 这个词不是修饰神经网络的。即通过神经网络来增强。

这样理解的话，句子里的"generating"就没有翻译出来了。我觉得后半句是一个系表结构，“augmented ..."作为一个定语从句来修饰生成的神经网络。

我认为原翻译理解是对的，就是少了一个”不“，即”在这里我们不再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强“

linhe0x0 · 2017-07-25T05:30:30Z

@MoutainOne 校对完了吗？

linhe0x0 · 2017-07-25T05:30:42Z

@sunshine940326 别忘了来校对啊

MountainOne · 2017-07-25T06:32:23Z

今天可以校对完@sqrthree

sunshine940326 · 2017-07-25T03:31:59Z

TODO/the-future-of-deep-learning.md

- New forms of learning that make the above possible—allowing models to move away from just differentiable transforms.
- Models that require less involvement from human engineers—it shouldn't be your job to tune knobs endlessly.
- Greater, systematic reuse of previously learned features and architectures; meta-learning systems based on reusable and modular program subroutines.
+- 在比目前可比较的层次更丰富的原始数据之上，构建更贴近通用计算机程序的模型 —— 也就是说我们如何能够得到（数据的）**推理 (reasoning)**和**抽象(abstraction)**，这也是目前模型的根本弱点。


个人比较认同 MoutainOne 的想法

sunshine940326 · 2017-07-25T03:39:30Z

TODO/the-future-of-deep-learning.md

- Models that require less involvement from human engineers—it shouldn't be your job to tune knobs endlessly.
- Greater, systematic reuse of previously learned features and architectures; meta-learning systems based on reusable and modular program subroutines.
+- 在比目前可比较的层次更丰富的原始数据之上，构建更贴近通用计算机程序的模型 —— 也就是说我们如何能够得到（数据的）**推理 (reasoning)**和**抽象(abstraction)**，这也是目前模型的根本弱点。
+- 能够学习新的形式，使上述成为可能 —— 允许模型脱离（每一步之间的）微分变换 (differentiable transformation) 的限制。（译者注：神经网络的每一步传播本质上是一个微分变换）


允许模型摆脱积分变换的新的学习模式使得实现上述的模型成为可能。这样翻译读起来更加顺

sunshine940326 · 2017-07-25T05:52:09Z

TODO/the-future-of-deep-learning.md


-Additionally, do note that these considerations are not specific to the sort of supervised learning that has been the bread and butter of deep learning so far—rather, they are applicable to any form of machine learning, including unsupervised, self-supervised, and reinforcement learning. It is not fundamentally important where your labels come from or what your training loop looks like; these different branches of machine learning are just different facets of a same construct.
+此外，请注意，这些考虑并不只是特定于已经作为深度学习基础设施的有监督学习，而是适用于任何形式的机器学习，包括无监督、自监督及强化学习。你（训练数据）标签的来源或你的训练循环怎么样其实并不重要，机器学习的这些不同的分支只是同一结构的不同面而已。


1、此外，请注意，这些考虑并不只是特定于已经作为深度学习基础设施的有监督学习.... -> 此外，请注意，这些考虑并不是（很小的一个建议，去掉“只”）
1、你（训练数据）标签的来源 -> 训练数据标签的来源

MountainOne · 2017-07-25T06:47:20Z

TODO/the-future-of-deep-learning.md


-Another important AutoML direction is to learn model architecture jointly with model weights. Because training a new model from scratch every time we try a slightly different architecture is tremendously inefficient, a truly powerful AutoML system would manage to evolve architectures at the same time as the features of the model are being tuned via backprop on the training data, thus eliminating all computational redundancy. Such approaches are already starting to emerge as I am writing these lines.
+另一个重要的AutoML 方向是与模型权重一起学习模型架构。因为每次尝试一个稍微不同的架构时，都会从头开始训练一个新的模型，所以一个真正强大的 AutoML 系统将通过对训练数据的反馈来调整模型的特征，同时管理网络架构，进而消除所有计算冗余。这样的方法已经开始出现，因为我正在写这些东西。


因为每次尝试一个稍微不同的架构时，都会从头开始训练一个新的模型 =>
因为每次尝试一个稍微不同的架构都需要重新训练模型是非常低效的

"is tremendously inefficient"这句话没翻出来，但感觉也可不加，看译者取舍。

MountainOne · 2017-07-25T06:54:38Z

TODO/the-future-of-deep-learning.md


-If models get more complex and are built on top of richer algorithmic primitives, then this increased complexity will require higher reuse between tasks, rather than training a new model from scratch every time we have a new task or a new dataset. Indeed, a lot datasets would not contain enough information to develop a new complex model from scratch, and it will become necessary to leverage information coming from previously encountered datasets. Much like you don't learn English from scratch every time you open a new book—that would be impossible. Besides, training models from scratch on every new task is very inefficient due to the large overlap between the current tasks and previously encountered tasks.
+如果模型变得更加复杂，并且建立在更丰富的算法原语之上，那么这种增加的复杂性将需要更高的任务之间的复用，而不是每当我们有一个新的任务或一个新的数据集从头开始训练一个新的模型。实际上，很多数据集并没有包含足够的信息来从头开发新的复杂模型，而且利用来自先前遇到的数据集的信息也是有必要的。 这就像你每次打开新书时都不会从头开始学习英语 —— 这是不可能的。此外，由于当前任务与以前遇到的任务之间的重叠很大，对每个新任务训练模型的效率是非常低的。


对每个新任务训练模型的效率是非常低的。 =>
对每个新任务重头开始训练模型的效率是非常低的。

”from scratch"在该句中比较重要，还是翻译出来好些。

MountainOne · 2017-07-25T07:17:31Z

TODO/the-future-of-deep-learning.md

+  环境及其自身，由于其丰富的算法性质，它们能够获得更强的泛化能力。
+- 特别地，模型将混合提供正式推理、搜索和抽象能力的算法模块，和提供非正式的直觉和模式识别功能的几何模块。AlphaGo（一个需要大量手动软件工程和人造设计决策的系统）提供了一个早期的范例，说明了符号与几何 AI 之间进行混合可能的样子。
+- 它们将不再由人类工程师手工打造，自动成长并使用存储在可复用子程序的全局库中的模块化部件（通过在数千个前有任务和数据集上学习过的高性能模型而演变出现的库）。由于常见的问题解决模式是通过元学习系统来识别的，它们将变成可复用的子程序并被添加到全局库中，像极了当代软件工程中的函数和类，进而实现了抽象的能力。
+- 这个全局库和相关的模式增长系统将能够实现某种形式的人类「极端泛化」：给定一个新的任务、一个新的情况，该系统将能够使用非常少的数据组装适合于任务的新的有效模型。这归功于：第一，丰富的泛化良好且类似程序的的原语；第二，丰富的类似任务的经验。同样地，人类也可以用很少的游戏时间来学习复杂的新游戏，因为他们有许多以前的游戏的经验，并且从以前的经验得出的模型是抽象的和类似程序的，而不是刺激和行动之间的基本映射。


丰富的泛化良好且类似程序的的原语 =>
可以像原语一样使用，丰富且泛化良好的程序（包）

感觉这句话的主语是程序而不是原语，一个个的程序（网络）像原语一样按照规则组合成一个有效的模型，就像我们用编程语言写一个程序一样。应该是这个意思

MountainOne · 2017-07-25T07:23:53Z

@sqrthree @changkun 校对完成。译者翻译的很棒！

changkun · 2017-07-25T08:34:01Z

@MoutainOne 非常感谢宝贵的翻译经验, 已经根据校对意见修改完毕~

MountainOne · 2017-07-25T08:53:31Z

TODO/the-future-of-deep-learning.md


-At the most basic level, such a system would simply tune the number of layers in a stack, their order, and the number of units or filters in each layer. This is commonly done with libraries such as Hyperopt, which we discussed in Chapter 7 (Note: of [Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras&amp;a_bid=76564dff)). But we can also be far more ambitious, and attempt to learn an appropriate architecture from scratch, with as few constraints as possible. This is possible via reinforcement learning, for instance, or genetic algorithms.
+在最基本的级别上，这样的系统将简单地调整（网络）栈中的层数、它们的顺序以及每一层中的单元或过滤器的数量。 这通常可以由诸如 Hyperopt 的库来完成，我们在第 7 章中讨论过（注：[Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras& a_bid=76564dff)）。但是我们也可以更加有野心，尝试从头开始学习一个适当的网络架构，尽可能少的约束。这可以通过加强学习来实现，例如遗传算法。


注：[Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras& a_bid=76564dff)）

超链接未正确显示，建议调整下格式

sunshine940326 · 2017-07-25T06:50:41Z

TODO/the-future-of-deep-learning.md


-As we noted in our previous post, a necessary transformational development that we can expect in the field of machine learning is a move away from models that perform purely pattern recognition and can only achieve local generalization, towards models capable of abstraction and reasoning, that can achieve extreme generalization. Current AI programs that are capable of basic forms of reasoning are all hard-coded by human programmers: for instance, software that relies on search algorithms, graph manipulation, formal logic. In DeepMind's AlphaGo, for example, most of the "intelligence" on display is designed and hard-coded by expert programmers (e.g. Monte-Carlo tree search); learning from data only happens in specialized submodules (value networks and policy networks). But in the future, such AI systems may well be fully learned, with no human involvement.
+正如我们在上一篇文章中指出的那样，我们可以预计的是，机器学习领域开发的一个必要转型就是从使用模型本身进行纯模式识别并实现局部泛化当中脱离开来，转为能够进行抽象及推理的模型，从而实现极端泛化（extreme generalization）。目前的 AI 程序所具有的基本形式的推理能力均为程序员们手动写死的代码，例如：依赖搜索算法、图操作、形式逻辑的软件；又比如 DeepMind 的 AlphaGo，大多数所谓的「智能」其实都是由专业程序员写死实现的（例如 Monte-Carlo 树搜索）；从数据中学习只发生在特殊的一些子模块（价值网络及策略网络）中。但是，这样的 AI 系统在未来可能会在没有人为参与的情况下被充分学习。


1、进行纯模式识别并实现局部泛化当中脱离开来， -> 进行纯模式识别并仅仅只能实现局部泛化当中脱离开来，（only没有翻译出“仅仅，只有”的意思）
2、大多数所谓的「智能」其实都是由专业程序员写死实现的 -> 大多数所谓的「智能」其实都是被专业程序员设计并写死实现的（designed没有翻译出，并且这句我认为使用被动句更好一些）

sunshine940326 · 2017-07-25T08:01:40Z

TODO/the-future-of-deep-learning.md


-What could be the path to make this happen? Consider a well-known type of network: RNNs. Importantly, RNNs have slightly less limitations than feedforward networks. That is because RNNs are a bit more than a mere geometric transformation: they are geometric transformations repeatedly applied inside a for loop. The temporal for loop is itself hard-coded by human developers: it is a built-in assumption of the network. Naturally, RNNs are still extremely limited in what they can represent, primarily because each step they perform is still just a differentiable geometric transformation, and the way they carry information from step to step is via points in a continuous geometric space (state vectors). Now, imagine neural networks that would be "augmented" in a similar way with programming primitives such as for loops—but not just a single hard-coded for loop with a hard-coded geometric memory, rather, a large set of programming primitives that the model would be free to manipulate to expand its processing function, such as if branches, while statements, variable creation, disk storage for long-term memory, sorting operators, advanced datastructures like lists, graphs, and hashtables, and many more. The space of programs that such a network could represent would be far broader than what can be represented with current deep learning models, and some of these programs could achieve superior generalization power.
+什么可以使得这种情况成为可能呢？考虑一个众所周知的网络类型：RNN。很重要的一点就是，RNN 的局限性远小于前馈神经网。这是因为 RNN 不仅仅只是一个简单几何变换，而是在 for 循环里不断重复的几何变换。时间 for 循环本身由程序员编写，这是网络本身的假设。当然，RNN 在它们能够表示的方面依然十分有限，主要原因是它们执行的每个步骤都是一个可微的几何变换，并且它们每一步传递信息的方式是通过连续几何空间中的点（状态向量）。现在，想象神经网络将以类似编程原语（例如 for 循环）的方式「增强」—— 但不仅仅是一个单一的写死的具有写死的几何记忆的 for 循环，而是一组大量的编程原语，使得模型能够自由的操纵并且扩充它的处理函数，例如 if 条件分支、while 循环语句、变量创建、长期记忆的磁盘存储、排序运算符、诸如列表、图、哈希表等的高级数据结构等等。这样一个网络可以表示的程序的空间将远大于当前深度学习模型所能表达的范围，其中一些程序可以实现更高的泛化能力。


1、前馈神经网 -> 前馈神经网络（多一个字好像更好一点，捂脸，都是这些小问题）
2、时间 for 循环本身由程序员编写 -> 大多数 for 循环本身由程序员写死的（时间读起来怪怪的，hard-coded 没有翻译出来）
3、这是网络本身的假设 -> 这是网络内置的假设
4、现在，想象神经网络将以类似编程原语（例如 for 循环）的方式「增强」—— 但不仅仅是一个单一的写死的具有写死的几何记忆的 for 循环，-> 现在，想象神经网络将以类似编程原语（例如 for 循环，但不仅仅是一个单一的写死的具有写死的几何记忆的 for 循环，而是一组大量的编程原语）的方式「增强」。（感觉这样表达意思更清晰，）

「时间 for」是 RNN 特有的代码内容，改成「大多数 for」要不加一个译者注？

网络本身的假设、网络原本的假设似乎没啥问题，改成「内置的假设」有点怪.. 有点违背常见说法

1、「时间 for」如果是特有的内容可以加上引号之类的，开始读不理解”时间“是什么意思
2、built-in 翻译为”内置“的情况多一点，所以提了建议，你可以再斟酌斟酌

sunshine940326 · 2017-07-25T08:18:54Z

TODO/the-future-of-deep-learning.md


-In a word, we will move away from having on one hand "hard-coded algorithmic intelligence" (handcrafted software) and on the other hand "learned geometric intelligence" (deep learning). We will have instead a blend of formal algorithmic modules that provide reasoning and abstraction capabilities, and geometric modules that provide informal intuition and pattern recognition capabilities. The whole system would be learned with little or no human involvement.
+总而言之，我们将远离写死的算法智能（手工软件）和学会的几何智能（深度学习），取而代之的是去提供提供推理和抽象能力的正式算法模块和提供非正式直觉和模式识别功能的几何模块，整个系统的学习将很少或甚至没有人参与。


1、提供提供 -> 提供
2、取而代之的是去提供提供推理和抽象能力的正式算法模块和提供非正式直觉和模式识别功能的几何模块
3、a blend of 没有翻译到
4、整个系统的学习将很少或甚至没有人参与 -> 很少甚至没有人参与整个系统的学习（语序稍微调整了下，或和甚至可以取一个，两个连在一起感觉不很通顺）

sunshine940326 · 2017-07-25T09:04:46Z

TODO/the-future-of-deep-learning.md


-A related subfield of AI that I think may be about to take off in a big way is that of program synthesis, in particular neural program synthesis. Program synthesis consists in automatically generating simple programs, by using a search algorithm (possibly genetic search, as in genetic programming) to explore a large space of possible programs. The search stops when a program is found that matches the required specifications, often provided as a set of input-output pairs. As you can see, is it highly reminiscent of machine learning: given "training data" provided as input-output pairs, we find a "program" that matches inputs to outputs and can generalize to new inputs. The difference is that instead of learning parameter values in a hard-coded program (a neural network), we generate source code via a discrete search process.
+有一个相关的 AI 子领域我认为可能会出现巨大突破，那就是程序合成（Program Synthesis），尤其是神经程序合成（Neural Program Synthesis）。程序合成在于通过使用搜索算法（可能的遗传搜索、遗传编程）自动生成简单的程序，从而探索包含可能程序的一个更大的空间。当找到符合要求的程序后，停止搜索，并作为一组输入输入对来提供。正如你所看到的，是不是让人联想起机器学习：给定作为输入输出对的训练数据，找到一个程序使其匹配输入输出对，并能够泛化新的输入。不同之处在于，我们不用去学习写死程序的参数，而是通过离散的搜索过程来生成源代码。


1、从而探索包含可能程序的一个更大的空间 -> 从而探索可能程序的一个更大的空间
2、是不是让人联想起机器学习 -> 这让我们高度联想到机器学习（highly，高度的，作者强调的语气没有翻译出来）
3、给定作为输入输出对的训练数据 -> 给定”实验数据“作为输入输出对（建议去掉“对”，要不然就再调整下语句，容易产生歧义，“对的“和”输入输出对“）；
4、a neural network 没有翻译出来

sunshine940326 · 2017-07-25T09:13:21Z

TODO/the-future-of-deep-learning.md


-I would definitely expect this subfield to see a wave of renewed interest in the next few years. In particular, I would expect the emergence of a crossover subfield in-between deep learning and program synthesis, where we would not quite be generating programs in a general-purpose language, but rather, where we would be generating neural networks (geometric data processing flows) augmented with a rich set of algorithmic primitives, such as for loops—and many others. This should be far more tractable and useful than directly generating source code, and it would dramatically expand the scope of problems that can be solved with machine learning—the space of programs that we can generate automatically given appropriate training data. A blend of symbolic AI and geometric AI. Contemporary RNNs can be seen as a prehistoric ancestor to such hybrid algorithmic-geometric models.
+我相当期待这个子领域在未来的几年里掀起一股新浪潮。特别地，我期望深度学习和程序合成之间能够再出现一个交叉子领域，在这里我们再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强，比如 for 循环等等。这会比直接生成源代码要容易得多，而且他会大大的扩展机器学习可以解决问题的范围 —— 我们可以自动生成给定适当训练数据的程序空间。一个符号 AI 与几何 AI 的混合。当代的 RNN 可以看做是这种混合算法与几何模型的鼻祖。


我认为原翻译理解是对的，就是少了一个”不“，即”在这里我们不再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强“

sunshine940326 · 2017-07-25T09:15:27Z

TODO/the-future-of-deep-learning.md


-I would definitely expect this subfield to see a wave of renewed interest in the next few years. In particular, I would expect the emergence of a crossover subfield in-between deep learning and program synthesis, where we would not quite be generating programs in a general-purpose language, but rather, where we would be generating neural networks (geometric data processing flows) augmented with a rich set of algorithmic primitives, such as for loops—and many others. This should be far more tractable and useful than directly generating source code, and it would dramatically expand the scope of problems that can be solved with machine learning—the space of programs that we can generate automatically given appropriate training data. A blend of symbolic AI and geometric AI. Contemporary RNNs can be seen as a prehistoric ancestor to such hybrid algorithmic-geometric models.
+我相当期待这个子领域在未来的几年里掀起一股新浪潮。特别地，我期望深度学习和程序合成之间能够再出现一个交叉子领域，在这里我们再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强，比如 for 循环等等。这会比直接生成源代码要容易得多，而且他会大大的扩展机器学习可以解决问题的范围 —— 我们可以自动生成给定适当训练数据的程序空间。一个符号 AI 与几何 AI 的混合。当代的 RNN 可以看做是这种混合算法与几何模型的鼻祖。


1、这会比直接生成源代码要容易得多 -> 这会比直接生成源代码要容易和有用得多（useful没有翻译出）
2、训练数据 -> 测试数据
3、我们可以自动生成给定适当训练数据的程序空间 -> 我们可以自动生成给定适当测试数据的程序的范围（前面是范围，后面最好也使用范围吧，前后一致）

sunshine940326 · 2017-07-25T09:22:58Z

TODO/the-future-of-deep-learning.md


-I would definitely expect this subfield to see a wave of renewed interest in the next few years. In particular, I would expect the emergence of a crossover subfield in-between deep learning and program synthesis, where we would not quite be generating programs in a general-purpose language, but rather, where we would be generating neural networks (geometric data processing flows) augmented with a rich set of algorithmic primitives, such as for loops—and many others. This should be far more tractable and useful than directly generating source code, and it would dramatically expand the scope of problems that can be solved with machine learning—the space of programs that we can generate automatically given appropriate training data. A blend of symbolic AI and geometric AI. Contemporary RNNs can be seen as a prehistoric ancestor to such hybrid algorithmic-geometric models.
+我相当期待这个子领域能在未来的几年里掀起一股新浪潮。特别地，我期望深度学习和程序合成之间能够再出现一个交叉子领域，在这里我们不再用通用语言来写程序，而是生成通过丰富的算法原语集增强的神经网络（几何数据处理流），比如 for 循环等等。这会比直接生成源代码要容易得多，而且他会大大的扩展机器学习可以解决问题的范围 —— 我们可以自动生成给定适当训练数据的程序空间。一个符号 AI 与几何 AI 的混合。当代的 RNN 可以看做是这种混合算法与几何模型的鼻祖。


1、这会比直接生成源代码要容易得多 -> 这会比直接生成源代码要容易和有用得多（useful没有翻译出）
2、我们可以自动生成给定适当训练数据的程序空间 -> 我们可以自动生成给定适当测试数据的程序的范围（训练数据-> 测试数据，前面使用的是范围，后面最好也使用范围，前后一致）

前面用的 scope, 后面是 space. 应该不用保持一致吧? 此外 training data 是训练数据，不是测试数据 test data

1、sorry，scope 和 space看错了
2、training data 的你是对的，开始觉得训练数据这个有点怪，但确实是训练数据，和测试数据不同

changkun · 2017-07-25T10:22:37Z

@sunshine940326 已根据相关意见修改完毕~

sunshine940326 · 2017-07-25T13:07:56Z

TODO/the-future-of-deep-learning.md


-When this starts happening, the jobs of machine learning engineers will not disappear—rather, engineers will move higher up the value creation chain. They will start putting a lot more effort into crafting complex loss functions that truly reflect business goals, and understanding deeply how their models impact the digital ecosystems in which they are deployed (e.g. the users that consume the model's predictions and generate the model's training data) —problems that currently only the largest company can afford to consider.
+当这种情况开始发生时，机器学习工程师的工作并不会消失 —— 相反，工程师将在价值创造链上站的更高。他们将开始更多地努力制定真正反映业务目标的复杂损失函数，并更加深入了解他们的模型如何影响其部署的数字生态系统（例如，消耗模型预测内容并生成模型训练数据的用户） —— 考虑那些目前只有大公司才能考虑的问题。


一个特别小的细节，破折号前面是中文标应该是不用加空格的

sunshine940326 · 2017-07-25T13:15:30Z

TODO/the-future-of-deep-learning.md


-If machine learning models become more like programs, then they will mostly no longer be differentiable—certainly, these programs will still leverage continuous geometric layers as subroutines, which will be differentiable, but the model as a whole would not be. As a result, using backpropagation to adjust weight values in a fixed, hard-coded network, cannot be the method of choice for training models in the future—at least, it cannot be the whole story. We need to figure out to train non-differentiable systems efficiently. Current approaches include genetic algorithms, "evolution strategies", certain reinforcement learning methods, and ADMM (alternating direction method of multipliers). Naturally, gradient descent is not going anywhere—gradient information will always be useful for optimizing differentiable parametric functions. But our models will certainly become increasingly more ambitious than mere differentiable parametric functions, and thus their automatic development (the "learning" in "machine learning") will require more than backpropagation.
+如果机器学习模型变得更像程序，那么它们将几乎不再是可微的 —— 当然，这些程序依然会将连续的几何图层作为可微的子程序，但是整个模型却不会这样。因此，使用反向传播来调整固定、写死的网络权重不能成为未来训练模型的首选方法 —— 至少不能是唯一的方法。我们需要找出有效地训练不可微系统的方法。目前的方法包括遗传算法、「进化策略」、某些强化学习方法和 ADMM（乘子交替方向法）。自然地，梯度下降不会被淘汰 —— 因为梯度信息总是对优化可微参数的函数有用。但是，我们的模型肯定会变得越来越有野心，而不仅仅只满足于可微参数的函数。因此它们的自动开发（「机器学习」中的「学习」）将需要的不仅仅只普通的反向传播。


1、我们的模型肯定会变得越来越有野心，而不仅仅只满足于可微参数的函数。 -> 和可微参数的函数相比我们的模型肯定会变得越来越有野心。（但从英文来看，than前后应该是同级比较吧，翻译的感觉不是同级的意思）

感觉原翻译更恰当

仍更倾向于原译法

changkun · 2017-07-25T21:14:41Z

@sunshine940326 已根据校对意见修改完毕, 校对非常细心, 感谢~

@sqrthree 本文校对修改完毕

linhe0x0

还有一丢丢小问题麻烦调整下就可以 merge 了哈

linhe0x0 · 2017-07-26T05:17:05Z

TODO/the-future-of-deep-learning.md

-You can read the first part here: [The Limitations of Deep Learning](https://blog.keras.io/the-limitations-of-deep-learning.html).
+这篇文章改编自我的书 [Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras&a_bid=76564dff)（Manning 出版社）第 9 章第 3 节（译者注：「结论」一章最后一小节）。
+它是讨论当前深度学习的局限性及其未来系列文章的第二篇。
+你可以在这里阅读第一篇：[深度学习的局限性](https://blog.keras.io/the-limitations-of-deep-learning.html)。


第一篇在这：https://github.com/xitu/gold-miner/blob/master/TODO/the-limitations-of-deep-learning.md

changkun · 2017-07-26T05:28:20Z

@sqrthree 已经改好~

linhe0x0 · 2017-07-26T06:15:09Z

已经 merge 啦~ 快快麻溜发布到掘金专栏然后给我发下链接，方便及时添加积分哟。

changkun · 2017-07-27T05:01:48Z

@sqrthree 掘金链接: https://juejin.im/post/597843506fb9a06ba4747db5 辛苦啦~

(add): 增加初稿

d74b061

linhe0x0 added the 校对认领 label Jul 22, 2017

linhe0x0 added the 正在校对 label Jul 23, 2017

linhe0x0 removed the 校对认领 label Jul 23, 2017

MountainOne reviewed Jul 23, 2017

View reviewed changes

linhe0x0 mentioned this pull request Jul 24, 2017

深度学习的未来 #1930

Closed

linhe0x0 added the AI label Jul 24, 2017

sunshine940326 reviewed Jul 25, 2017

View reviewed changes

MountainOne reviewed Jul 25, 2017

View reviewed changes

(update): 根据校对意见进行修改

7d04294

MountainOne reviewed Jul 25, 2017

View reviewed changes

sunshine940326 reviewed Jul 25, 2017

View reviewed changes

(update): 根据校对意见进行修改2

3bf739e

sunshine940326 reviewed Jul 25, 2017

View reviewed changes

(update): 根据校对意见进行修改3

044c51c

linhe0x0 requested changes Jul 26, 2017

View reviewed changes

(update): 修改链接

5e7a4b3

linhe0x0 approved these changes Jul 26, 2017

View reviewed changes

linhe0x0 merged commit 3030659 into xitu:master Jul 26, 2017

linhe0x0 added 翻译完成 and removed 正在校对 labels Jul 26, 2017

changkun deleted the translate branch August 2, 2017 19:55

cdadar pushed a commit to cdadar/gold-miner that referenced this pull request Dec 8, 2017

深度学习的未来 (xitu#1941)

69592e6


		I would definitely expect this subfield to see a wave of renewed interest in the next few years. In particular, I would expect the emergence of a crossover subfield in-between deep learning and program synthesis, where we would not quite be generating programs in a general-purpose language, but rather, where we would be generating neural networks (geometric data processing flows) augmented with a rich set of algorithmic primitives, such as for loops—and many others. This should be far more tractable and useful than directly generating source code, and it would dramatically expand the scope of problems that can be solved with machine learning—the space of programs that we can generate automatically given appropriate training data. A blend of symbolic AI and geometric AI. Contemporary RNNs can be seen as a prehistoric ancestor to such hybrid algorithmic-geometric models.
		我相当期待这个子领域在未来的几年里掀起一股新浪潮。特别地，我期望深度学习和程序合成之间能够再出现一个交叉子领域，在这里我们再用语言来生成程序，而是通过具有丰富算法原语的神经网络（几何数据处理流）来增强，比如 for 循环等等。这会比直接生成源代码要容易得多，而且他会大大的扩展机器学习可以解决问题的范围 —— 我们可以自动生成给定适当训练数据的程序空间。一个符号 AI 与几何 AI 的混合。当代的 RNN 可以看做是这种混合算法与几何模型的鼻祖。


		Additionally, do note that these considerations are not specific to the sort of supervised learning that has been the bread and butter of deep learning so far—rather, they are applicable to any form of machine learning, including unsupervised, self-supervised, and reinforcement learning. It is not fundamentally important where your labels come from or what your training loop looks like; these different branches of machine learning are just different facets of a same construct.
		此外，请注意，这些考虑并不只是特定于已经作为深度学习基础设施的有监督学习，而是适用于任何形式的机器学习，包括无监督、自监督及强化学习。你（训练数据）标签的来源或你的训练循环怎么样其实并不重要，机器学习的这些不同的分支只是同一结构的不同面而已。


		Another important AutoML direction is to learn model architecture jointly with model weights. Because training a new model from scratch every time we try a slightly different architecture is tremendously inefficient, a truly powerful AutoML system would manage to evolve architectures at the same time as the features of the model are being tuned via backprop on the training data, thus eliminating all computational redundancy. Such approaches are already starting to emerge as I am writing these lines.
		另一个重要的AutoML 方向是与模型权重一起学习模型架构。因为每次尝试一个稍微不同的架构时，都会从头开始训练一个新的模型，所以一个真正强大的 AutoML 系统将通过对训练数据的反馈来调整模型的特征，同时管理网络架构，进而消除所有计算冗余。这样的方法已经开始出现，因为我正在写这些东西。


		If models get more complex and are built on top of richer algorithmic primitives, then this increased complexity will require higher reuse between tasks, rather than training a new model from scratch every time we have a new task or a new dataset. Indeed, a lot datasets would not contain enough information to develop a new complex model from scratch, and it will become necessary to leverage information coming from previously encountered datasets. Much like you don't learn English from scratch every time you open a new book—that would be impossible. Besides, training models from scratch on every new task is very inefficient due to the large overlap between the current tasks and previously encountered tasks.
		如果模型变得更加复杂，并且建立在更丰富的算法原语之上，那么这种增加的复杂性将需要更高的任务之间的复用，而不是每当我们有一个新的任务或一个新的数据集从头开始训练一个新的模型。实际上，很多数据集并没有包含足够的信息来从头开发新的复杂模型，而且利用来自先前遇到的数据集的信息也是有必要的。这就像你每次打开新书时都不会从头开始学习英语 —— 这是不可能的。此外，由于当前任务与以前遇到的任务之间的重叠很大，对每个新任务训练模型的效率是非常低的。


		At the most basic level, such a system would simply tune the number of layers in a stack, their order, and the number of units or filters in each layer. This is commonly done with libraries such as Hyperopt, which we discussed in Chapter 7 (Note: of [Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras&a_bid=76564dff)). But we can also be far more ambitious, and attempt to learn an appropriate architecture from scratch, with as few constraints as possible. This is possible via reinforcement learning, for instance, or genetic algorithms.
		在最基本的级别上，这样的系统将简单地调整（网络）栈中的层数、它们的顺序以及每一层中的单元或过滤器的数量。这通常可以由诸如 Hyperopt 的库来完成，我们在第 7 章中讨论过（注：[Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras& a_bid=76564dff)）。但是我们也可以更加有野心，尝试从头开始学习一个适当的网络架构，尽可能少的约束。这可以通过加强学习来实现，例如遗传算法。


		As we noted in our previous post, a necessary transformational development that we can expect in the field of machine learning is a move away from models that perform purely pattern recognition and can only achieve local generalization, towards models capable of abstraction and reasoning, that can achieve extreme generalization. Current AI programs that are capable of basic forms of reasoning are all hard-coded by human programmers: for instance, software that relies on search algorithms, graph manipulation, formal logic. In DeepMind's AlphaGo, for example, most of the "intelligence" on display is designed and hard-coded by expert programmers (e.g. Monte-Carlo tree search); learning from data only happens in specialized submodules (value networks and policy networks). But in the future, such AI systems may well be fully learned, with no human involvement.
		正如我们在上一篇文章中指出的那样，我们可以预计的是，机器学习领域开发的一个必要转型就是从使用模型本身进行纯模式识别并实现局部泛化当中脱离开来，转为能够进行抽象及推理的模型，从而实现极端泛化（extreme generalization）。目前的 AI 程序所具有的基本形式的推理能力均为程序员们手动写死的代码，例如：依赖搜索算法、图操作、形式逻辑的软件；又比如 DeepMind 的 AlphaGo，大多数所谓的「智能」其实都是由专业程序员写死实现的（例如 Monte-Carlo 树搜索）；从数据中学习只发生在特殊的一些子模块（价值网络及策略网络）中。但是，这样的 AI 系统在未来可能会在没有人为参与的情况下被充分学习。


		What could be the path to make this happen? Consider a well-known type of network: RNNs. Importantly, RNNs have slightly less limitations than feedforward networks. That is because RNNs are a bit more than a mere geometric transformation: they are geometric transformations repeatedly applied inside a for loop. The temporal for loop is itself hard-coded by human developers: it is a built-in assumption of the network. Naturally, RNNs are still extremely limited in what they can represent, primarily because each step they perform is still just a differentiable geometric transformation, and the way they carry information from step to step is via points in a continuous geometric space (state vectors). Now, imagine neural networks that would be "augmented" in a similar way with programming primitives such as for loops—but not just a single hard-coded for loop with a hard-coded geometric memory, rather, a large set of programming primitives that the model would be free to manipulate to expand its processing function, such as if branches, while statements, variable creation, disk storage for long-term memory, sorting operators, advanced datastructures like lists, graphs, and hashtables, and many more. The space of programs that such a network could represent would be far broader than what can be represented with current deep learning models, and some of these programs could achieve superior generalization power.
		什么可以使得这种情况成为可能呢？考虑一个众所周知的网络类型：RNN。很重要的一点就是，RNN 的局限性远小于前馈神经网。这是因为 RNN 不仅仅只是一个简单几何变换，而是在 for 循环里不断重复的几何变换。时间 for 循环本身由程序员编写，这是网络本身的假设。当然，RNN 在它们能够表示的方面依然十分有限，主要原因是它们执行的每个步骤都是一个可微的几何变换，并且它们每一步传递信息的方式是通过连续几何空间中的点（状态向量）。现在，想象神经网络将以类似编程原语（例如 for 循环）的方式「增强」—— 但不仅仅是一个单一的写死的具有写死的几何记忆的 for 循环，而是一组大量的编程原语，使得模型能够自由的操纵并且扩充它的处理函数，例如 if 条件分支、while 循环语句、变量创建、长期记忆的磁盘存储、排序运算符、诸如列表、图、哈希表等的高级数据结构等等。这样一个网络可以表示的程序的空间将远大于当前深度学习模型所能表达的范围，其中一些程序可以实现更高的泛化能力。


		In a word, we will move away from having on one hand "hard-coded algorithmic intelligence" (handcrafted software) and on the other hand "learned geometric intelligence" (deep learning). We will have instead a blend of formal algorithmic modules that provide reasoning and abstraction capabilities, and geometric modules that provide informal intuition and pattern recognition capabilities. The whole system would be learned with little or no human involvement.
		总而言之，我们将远离写死的算法智能（手工软件）和学会的几何智能（深度学习），取而代之的是去提供提供推理和抽象能力的正式算法模块和提供非正式直觉和模式识别功能的几何模块，整个系统的学习将很少或甚至没有人参与。


		A related subfield of AI that I think may be about to take off in a big way is that of program synthesis, in particular neural program synthesis. Program synthesis consists in automatically generating simple programs, by using a search algorithm (possibly genetic search, as in genetic programming) to explore a large space of possible programs. The search stops when a program is found that matches the required specifications, often provided as a set of input-output pairs. As you can see, is it highly reminiscent of machine learning: given "training data" provided as input-output pairs, we find a "program" that matches inputs to outputs and can generalize to new inputs. The difference is that instead of learning parameter values in a hard-coded program (a neural network), we generate source code via a discrete search process.
		有一个相关的 AI 子领域我认为可能会出现巨大突破，那就是程序合成（Program Synthesis），尤其是神经程序合成（Neural Program Synthesis）。程序合成在于通过使用搜索算法（可能的遗传搜索、遗传编程）自动生成简单的程序，从而探索包含可能程序的一个更大的空间。当找到符合要求的程序后，停止搜索，并作为一组输入输入对来提供。正如你所看到的，是不是让人联想起机器学习：给定作为输入输出对的训练数据，找到一个程序使其匹配输入输出对，并能够泛化新的输入。不同之处在于，我们不用去学习写死程序的参数，而是通过离散的搜索过程来生成源代码。


		I would definitely expect this subfield to see a wave of renewed interest in the next few years. In particular, I would expect the emergence of a crossover subfield in-between deep learning and program synthesis, where we would not quite be generating programs in a general-purpose language, but rather, where we would be generating neural networks (geometric data processing flows) augmented with a rich set of algorithmic primitives, such as for loops—and many others. This should be far more tractable and useful than directly generating source code, and it would dramatically expand the scope of problems that can be solved with machine learning—the space of programs that we can generate automatically given appropriate training data. A blend of symbolic AI and geometric AI. Contemporary RNNs can be seen as a prehistoric ancestor to such hybrid algorithmic-geometric models.
		我相当期待这个子领域能在未来的几年里掀起一股新浪潮。特别地，我期望深度学习和程序合成之间能够再出现一个交叉子领域，在这里我们不再用通用语言来写程序，而是生成通过丰富的算法原语集增强的神经网络（几何数据处理流），比如 for 循环等等。这会比直接生成源代码要容易得多，而且他会大大的扩展机器学习可以解决问题的范围 —— 我们可以自动生成给定适当训练数据的程序空间。一个符号 AI 与几何 AI 的混合。当代的 RNN 可以看做是这种混合算法与几何模型的鼻祖。


		When this starts happening, the jobs of machine learning engineers will not disappear—rather, engineers will move higher up the value creation chain. They will start putting a lot more effort into crafting complex loss functions that truly reflect business goals, and understanding deeply how their models impact the digital ecosystems in which they are deployed (e.g. the users that consume the model's predictions and generate the model's training data) —problems that currently only the largest company can afford to consider.
		当这种情况开始发生时，机器学习工程师的工作并不会消失 —— 相反，工程师将在价值创造链上站的更高。他们将开始更多地努力制定真正反映业务目标的复杂损失函数，并更加深入了解他们的模型如何影响其部署的数字生态系统（例如，消耗模型预测内容并生成模型训练数据的用户） —— 考虑那些目前只有大公司才能考虑的问题。


		If machine learning models become more like programs, then they will mostly no longer be differentiable—certainly, these programs will still leverage continuous geometric layers as subroutines, which will be differentiable, but the model as a whole would not be. As a result, using backpropagation to adjust weight values in a fixed, hard-coded network, cannot be the method of choice for training models in the future—at least, it cannot be the whole story. We need to figure out to train non-differentiable systems efficiently. Current approaches include genetic algorithms, "evolution strategies", certain reinforcement learning methods, and ADMM (alternating direction method of multipliers). Naturally, gradient descent is not going anywhere—gradient information will always be useful for optimizing differentiable parametric functions. But our models will certainly become increasingly more ambitious than mere differentiable parametric functions, and thus their automatic development (the "learning" in "machine learning") will require more than backpropagation.
		如果机器学习模型变得更像程序，那么它们将几乎不再是可微的 —— 当然，这些程序依然会将连续的几何图层作为可微的子程序，但是整个模型却不会这样。因此，使用反向传播来调整固定、写死的网络权重不能成为未来训练模型的首选方法 —— 至少不能是唯一的方法。我们需要找出有效地训练不可微系统的方法。目前的方法包括遗传算法、「进化策略」、某些强化学习方法和 ADMM（乘子交替方向法）。自然地，梯度下降不会被淘汰 —— 因为梯度信息总是对优化可微参数的函数有用。但是，我们的模型肯定会变得越来越有野心，而不仅仅只满足于可微参数的函数。因此它们的自动开发（「机器学习」中的「学习」）将需要的不仅仅只普通的反向传播。

深度学习的未来 #1941

深度学习的未来 #1941

Conversation

changkun commented Jul 22, 2017

sunshine940326 commented Jul 23, 2017

linhe0x0 commented Jul 23, 2017

MountainOne commented Jul 23, 2017

linhe0x0 commented Jul 23, 2017

Choose a reason for hiding this comment

MountainOne Jul 23, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

changkun Jul 23, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

linhe0x0 commented Jul 25, 2017

linhe0x0 commented Jul 25, 2017

MountainOne commented Jul 25, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

MountainOne commented Jul 25, 2017

changkun commented Jul 25, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

changkun commented Jul 25, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

changkun Jul 25, 2017 • edited Loading

Choose a reason for hiding this comment

changkun commented Jul 25, 2017

linhe0x0 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

changkun commented Jul 26, 2017

linhe0x0 commented Jul 26, 2017

changkun commented Jul 27, 2017

MountainOne Jul 23, 2017 •

edited

Loading

changkun Jul 23, 2017 •

edited

Loading

changkun Jul 25, 2017 •

edited

Loading