近期风靡互联网的Deep Dream人工智能图像识别软件_风景图像识别软件

code: https://github.com/google/deepdream/blob/master/dream.ipynb

声明：本译文包含六篇以上原外文内容，现汇合一处，附有大量的链接。实际上，我本可以将其拆分成多篇短文陆续发表，但我不想那样做。大翻确实伤身，这样的文章又非常晦涩，所以用了好多天时间才完成。而长时间盯着电脑屏幕阅读、查阅、码字、审阅、修改搞得我身心憔悴。所以，原则上以后不再大翻。因个人水平有限，如文中存在某些翻译不当之处，请各位指正。欢迎转载本文，但请注明译文作者（我）和文章出处，尊重他人的辛苦劳动成果是美德。龙腾网 http://www.ltaaa.com

##1.1 第一节 《连线杂志》报道

These Google "Deep Dream" images are weirdly mesmerising
这些 Google "Deep Dream" 式图像很是奇妙

http://www.wired.co.uk/news/archive/2015-07/03/google-deep-dream

Technology / 03 July 15, by Daniel Culpan
科技版 | 2015-7-15，Daniel Culpan

Last month, Google revealed that it uses its own artificial intelligence program, known as Artificial Neural Networks, to classify and sort its images.
上个月，Google 展示了自有人工智能程序——这种被称作“人工神经网络”的东西，（可以）用来分类和整理其图像库。

The technology basically works by spotting patterns in pictures in order to identify them -- and it's already being used in Google's new photos app to recognise faces and animals.
这种技术主要以指认其中图案的方式来识别图片，而且它已经被应用于 Google 最新的照片管理软件（似乎是 Google I/O 2015 推出的 Google Photos）中，以识认面部和动物。

However, a new website called Deep Neural Net Dreams, created by Zain Shah, allows you to take your own photos and run them through Google's AI code. The results? Some seriously psychedelic and eye-popping images, from trippy landscapes and abstract cats to swirly recreations of famous paintings.
不管怎样，（现在已经有了）一个由 Zain Shah 建立的叫做“深度神经网络梦想”的新网站，它允许你上传图片然后通过 Google 的人工智能代码来处理。结果 （出图效果）怎样？一些图像很是迷幻又奇特，从奇幻的风景、抽象的猫咪到名画的旋涡状再创作，应有尽有。

You can upload your own photos and share them using #deepdream.
你可以上传自己的照片，然后加上 #deepdream 标签分享出去。

（以下为编辑选择的 Twitter 热门图片）








译者注：Ernst Haeckel，1834-1919，德国生物学家、医生。他也是优生学的先驱，但其一些理论和主张常体现出德国民族沙文主义，后被纳粹理论家利用，成为种族主义和社会达尔文主义的理由。此蜂鸟图片的原图可能出自（我不确定，因为我还没拜读过）其科学著作《自然的艺术形式》（即 Kunstformen der Natur 或 Art Forms in Nature），此书编排精巧、绘画精美绝伦，完美展现了自然界中诸多奇妙的生物和各种形式的对称美。


译者注：
Bhuddhabrot：佛像分形，这个词的来源是 Buddha（佛）与 Mandelbrot（曼德布罗特），属于数学分形理论的范畴；
Inceptionism：这是 Deep Dream 项目开源发布之前的名称，现已弃用。恕我不予翻译（难以选择用词，翻译会变味），下面的第三篇技术文章将对此着重介绍。

---

##1.2 第二节 《独立报》文章

###1.2.1 第二节正文：

Google has set its terrifying, dreaming image robots on the public
Google 释出了它那可怕、猜想图像的机器人

http://www.independent.co.uk/life-style/gadgets-and-tech/news/google-has-set-its-terrifying-dreaming-image-robots-on-the-public-10361298.html

Google said that the technology might allow us to understand where human creativity comes from — and now they're putting it to the test
前言：Google 说这项技术或能使我们理解人类创造力的来源，现在他们正将其推出（面向公众）进行测试。

Andrew Griffin | Biography | Thursday 02 July 2015
Andrew Griffin | 生物 | 2015-7-02，周四

Google has opened up its image recognising robots to anybody — letting people create strange, horrifying images out of their own pictures.
Google 已经向世人开放了其图像识别机器人，它能让人们基于自己的图片创作出奇怪、恐怖的（新）图像。

Google released the half-horrifying, half-amazing pictures that it had created itself last week, with pictures including a knight made of dogs. Now the company has made the “Deep Dream” software available on code-sharing website Github, where anyone can download it and run their own pictures through it.
Google 发布了一些其上周就已创作出来的既恐怖又神奇的图片，其中包含一幅由狗狗组成的骑士像。该公司已经将这个名为“Deep Dream”的软件项目放到了代码分享网站 GitHub 上开源，现在任何人都可以下载它运行并藉此处理自己的图像。

The software works by turning the image recognising computers on themselves. By telling the systems to over-interpret images, they would pick out otherwise meaningless things and exaggerate them — turning clouds into bizarre llamas, for instance.
这款软件是以计算机自身来识别图像的。让系统过分解读图像，然后就可以筛选出其他无含义的东西并夸大，比如把云朵转化为羊驼 （美洲驼）。

As with Google’s own images, the pictures tend to transform thing into animals — with dogs being a particular favourite — and eyes. They also tend to overlay everything with a swirly rainbow colouring.
与 Google 发布的图片一样，（软件的）出图效果趋向于把很多东西转换成动物，其中特别爱抢镜的是狗狗......还有眼睛，也往往以一种旋涡状的彩虹色把东西给覆盖住。





Users can download the software from Google's blogpost. Once it's all set up, users can feed in an image, choose which parts of the network should be amplified and how dramatically, and then see the trippy picture come out the other side.
用户可以从 Google 的博客站下载这款软件。当一切都准备就绪了，用户就能喂入一幅图像，选择网格线中的某部分进行增强 （夸张）、程度要多深，接着就能在另一侧看到效果图。

"It'll be interesting to see what imagery people are able to generate," wrote the Google engineers in the post introducing the tool. "If you post images to Google+, Facebook, or Twitter, be sure to tag them with #deepdream so other researchers can check them out too."
“看看人们能做出何种图像会很有意思”，Google 工程师在介绍此工具的博文中写道，“如果你要在 Google+、非死不可或推特上发布图片，记得加上 #deepdream 标签。如此，其他人就也能切克闹 （来看）一下。”

###1.2.2 第二节评论（Comments）：

VED from Victoria Institutions
It is just a proof that human brain has a software background. In fact, even the human body and possibly reality itself is designed and maintained by a supernatural software. Read: Codes of reality! What is language?
这正好表明人脑有一个（隐藏在）后台的软件。事实上，甚至连人类的躯体、很可能还有现实（世界）本身都是被一个超自然软件设计和维护的。那就读读这本书《现实的代码！是以何种语言写就的？》 （晕，难道是来卖书的？）

AA  -1
These images depend on an algorithym processing that infirmation in a certain way preprogrammed into the code. There is nothing mechanical about it as the code was written by humans. This is probably why certain patterns appear.
这些图像依赖于以预先编译好的某一方式来处理信息的算法。根本就没有什么机械式的东西，因为代码都是由人写的。这可能就是某些图形（频繁）出现的原因。

    Hug
    Yes. It is a program or algorithm created by people, programmers... The pictures see people, machine and the computer program may not realize it
    对。它是由人——程序猿创造的一个程序或算法......这些图片能看人，机器和计算机程序多半还无法理解。

         Hug
        the images can seen by people, not artificial intelligence...
        这些图像能被人理解 （译注：此人在纠正上一句的语法错误），而人工智能办不到......



##2.1 Hopes&Fears 网站评论文章

Is Google's Deep Dream art?
Google 的 Deep Dream 是艺术吗？

http://www.hopesandfears.com/hopes/culture/is-this-art/215039-deep-dream-google-art

Marina Galperina, July 14

A little while ago, Google set an Artificial Neural Network wild to “dream” on the internet, digging through visual data, “enhancing” parts of images and building on features it “recognizes” within by using its own datasets. The result was a hazy, swirling glaze of colorful noise and objects mercilessly shapeshifted over and over. Trippy! It “saw” a lot of dogs. (There are a lot of dogs on the internet.)
就在不久前，Google 把一个人工神经网络项目放到了互联网上“做梦”，它可以挖掘可视的数据，“增强”图像中某些部分，而且其特性是依靠自己的数据集来“识别”里面（的内容）。（出图）效果是朦胧的、旋涡状有噪点的彩釉色，（里面的）物体无情地反复变化。实在太迷幻了！它“看出”了很多狗 （网上有许多狗-_-||）。

But is this art?
不过，这是艺术吗？

Not only is the neural network “dreaming”, it's possibly making art. Or is it? Well, at least now that the “inceptionism” algorithm is freely available online, people have certainly been using it creatively. ( Fear and Loathing, Quake, donuts, the universe, sheep...) But as a project, is Google's #deepdream robot an art piece in its own right? Does the very fact that it produces images that are pleasing to humans make the algorithm (or, at least its corporate creator) a sort of an artist? Can you program artistic agency into a bot or an “AI,” or is that idea a logical fallacy? We digress.
神经网络不仅仅是在“梦想”，它可能（也）在进行艺术创作。或许是这样的？好吧，至少现在“inceptionism”算法可以在线随意获取，人们肯定已经在创造性地使用它了。（ 电影《恐惧拉斯维加斯》、 游戏《雷神之锤》、甜甜圈、宇宙、羊咩咩......）但作为一个项目，Google 的 Deep Dream 机器人本身是不是件艺术品？它确实能生成可以取悦人们的图像就使得这种算法（或者，至少它的作者）有点像一个艺术家？你能把艺术机构编程进一个机器人或“人工智能”（的东西）吗？或者那种想法就是一个逻辑谬误？（好吧，）我们跑题了。



“Inceptionism” algorithm as art, much like those #deepdream porn-bug-dicks (NSFW), is a twisted, confusing concept. For our inaugural Is this art? inquiry, we asked art critics, artists and artistic technology experts to share their thoughts.
“Inceptionism”算法作为艺术，很像那些（用）Deep Dream（创作的） 色情的牛B的把儿（不适合上班时间浏览 ，即少儿不宜），是一种扭曲、令人困惑的概念。言归正传，这算不算艺术？（为了）一探究竟，我们邀请了艺术评议者、艺术家和艺术界技术专家来说出他们的看法。


Paddy Johnson
Art critic, founder and editor of Art F City
艺术评议者，Art F City 网站的创办者兼编辑

Nope. It's a tool, not the product, so calling it art would be a little like an artist raising their hand and declaring their paintbrush art because they were so happy with the way they used it lay paint on a canvas. Even the software engineers make this distinction on the Google blog:
不是。它只是一个工具，而非产品，因此称其为艺术有点像艺术家抬起手然后宣称他们的笔刷是艺术就因为他们非常满意在画布上涂抹的使用方式。即使是此软件的开发者也在 Google 博客 （网志）上指出了这一区别：

Two weeks ago we blogged about a visualization tool designed to help us understand how neural networks work and what each layer has learned. In addition to gaining some insight on how these networks carry out classification tasks, we found that this process also generated some beautiful art.
两周前我们发表了一篇博文，其中阐述了一个可视化（视觉）工具被设计用来帮助我们理解神经网络的工作原理以及每层学到了什么东西。除为了获悉这些（神经）网络是如何进行分类任务（作业）的，我们还发现这个过程也可以生成一些漂亮的艺术作品。

It's worth noting, though, that in the US corporations are considered people, so Google could deem their software art. If we take as a given that one person's the intent to make art, thus creates a work of art (regardless of its merit) it's certainly a possible future for Google. It sounds a bit like a sci-novel though, so I'm guessing we won't see the corporation do that any time soon.
但值得注意的是，似乎，在（众多的）美国公司中德高望重，因此谷歌本可以将他们的软件定义为艺术。如果我们考虑到一个人搞艺术的意图，那么对 Google 来说，创造一个艺术作品（不管其价值如何）肯定是其一个可能的未来。尽管这听起来有点像科幻小说，那么我猜我们将看不到该公司即刻去做（这事儿）。


Rich Oglesby
Creator and editor of Prosthetic Knowledge
Tumblr（轻博客） Prosthetic Knowledge （里面都是些原创的奇异视频和动图）专栏的创建者和编辑

I would say there is a good case to say that it is, dependent (as is usually the case with art) on how you 'frame' the idea or interpret the semiotics of process and results.
我想说有一个很好的例子可以用来谈谈此事，这取决于（照着通常与艺术相关的例子）你如何“构建”那个想法或解释其过程和结果的符号学。

It is easy to dismiss this as a modern form of psychedelia, but what this has made accessible to the public the idea of an artificial psychology of an artificial intelligence, and how it can connect to us: humans and computers are pattern-recognition entities processing immediate information. Ideas such as pareidolia, habituation and the 'Tetris Effect' come to mind, how our repetitive mental actions leave their impressions on us. If art has the purpose to illustrate the human condition, #deepdream has become a high profile entry point in this field.
人们很容易认为这是一种现代形式的迷幻剂而拒绝 （即先入为主），但是这对大众开放了一个人工智能（软件）的人工心理学，并展示了它是怎样与我们连接（关联）的：人类和计算机（同）是处理即时信息的模式 （图形）识别的实体。像幻想性视错觉、习惯化和“俄罗斯方块效应”这样的想法到了脑海里，（而）我们的重复性心理活动又是怎样给我们留下了他们的印象。如果艺术有为了说明人类的处境的目的，（那么）Deep Dream 已成为这个领域的翘楚 （直译为：一个引人注目的入口点）。

Releasing the source code has brought some fun experimentation for creative coders (not seen since the release of the Microsoft Kinect, which itself has been a modern staple for tech art). It could be considered a PR move from Google itself, based on the success of the original Inceptionism post, and we can quickly get tired of the formulations based on the dataset (the constant stream of images featuring 'puppyslugs' will eventually bore everyone), but the subject of Neural Network Art is developing and what is interesting is where it can go next. Examples include 'LSTM' by Sebastian Schmieg which is an app which generates new texts from the works of Ray Kurzweil. Creative coder Samim has produced several generative projects using neural networks to create machine-generated Obama political speeches and TED Talks, or how computer vision interprets pornography. Matthew Plummer Fernandez has created a bot which utilizes deep learning to interpret pieces of artworks, and posts its impressions onto a Tumblr blog.
开放源代码给富有创意的程序员带来了一些有趣的实验（自微软发布 Kinect 以来还未曾见到过，它本身已经成为一种科技艺术的现代主体）。它可以被视为谷歌本身的一个公关行为，是以最初 Inceptionism 博文的成功 （即引起社会的极大关注）为起点的。我们很快就会厌倦基于数据集（恒定的“群狗荟萃”特色图像流最终将使大家感到厌烦）的公式化设置，但是神经网络艺术的话题还在发酵中，以及有趣的是它接下来将何去何从。包括 Sebastian Schmieg （德国柏林艺术家）做的“长短时记忆”（Long Short Term Memory）实验在内的例子，（都只）是一个从 Ray Kurzweil （似乎是一个推销垃圾科学进行谋利的欺世盗名之徒）的作品中生成新文本的应用。才华横溢的程序员 Samim 运用神经网络已经做出了几个有生产力的项目，来创作机器生成版的 奥巴马政治演说和 TED 演讲，或者 计算机视觉应如何解译色情内容。Matthew Plummer Fernandez（也）创造了一个利用深度学习来解读艺术作品的 机器人，并把效果发表在 Tumblr 轻博客上。


Anthony Antonellis
Artist（艺术家）

Google Deep Dream is a medium. On its own it's not art, but the images it’s being used to create can be art. Fugly art.
Google 的 Deep Dream 项目（只）是个媒介。其本身并不是艺术，不过它出的图可以很艺术——那种丑爆了 （fucking ugly）的艺术。

It reminds me of the generative fractal computer art from in the 80s that filled up the columns of my grade school textbooks. Some of the results look like trippy scenes that could be used in a Pixar version of Fantasia. Conceptually it is very interesting; aesthetically it looks like visual Morgellons disease. I'm sure there could be some compelling results; it seems similar to the Photoshop Content-Aware Fill trend. I’m always happy to see a medium that so much of the public enjoys experimenting and playing with, I just wish those results would get printed out and put on their fridges rather than vomited onto my feed like the mid-2015 version of duck face.
它让我想起了八十年代里充斥在我的小学课本侧栏中的（批量）生成的不规则碎形计算机艺术。一些结果看起来像在皮克斯版的《幻想曲》中可能都运用过的迷幻场景。从概念上讲是很有趣的，以美学观点来看它像视觉莫吉隆斯症。 （译注：莫吉隆斯症是尚未被确认的“不治之症”，有些人坚信他们得了这种绝症，并抱团建立病患组织，但美国政府和主流医学界仍坚称这是精神疾病带来的妄想，由此亦引发了阴谋论。）我肯定会有一些引人注目的结果，似乎有类似于 Photoshop 中“内容感知填充工具”的趋势。我总是乐见一个能让如此多的公众乐于尝试并玩耍的媒介，我只是希望那些结果可以被打印出来，然后贴在他们（家）的冰箱上，而不是像今年年中版本的鸭子脸那样吐到我的（新闻）订阅源里。

Google Deep Dream is our punishment for not liking Google Glass.
Deep Dream 是 Google 对我们不喜欢 Google Glass （而做出）的惩罚。


Ben Davis
Art critic, author of 9.5 Theses on Art and Class, National Art Critic for artnet News
艺术评论家、《9.5 Theses on Art and Class》一书的作者、artnet 新闻网站的民族艺术评论家

Short answer: Of course it's art! There's no limit to what you can classify as "art." The question is only ever whether it's good art. And people seem to be very amused by it.
简短回答：当然是艺术啦！界定何为“艺术”并没有限制。问题只在于它是不是个好艺术。而且看起来人们都乐坏了。

Longer answer: It's not art.The way it is getting used is essentially like a psychedelic Instagram filter, and the results are actually a bit repetitious, don't you think? There were already very, very striking images produced by algorithmic means, so I'm not sure what the hoopla is.
啰嗦点的回答（吐槽：还真够哆嗦的，简直就是BS）：这不是艺术。它在运用的方法本质上类似一个迷幻的 Instagram（应用）滤镜，而且结果实际上有点重复，你不觉得吗？（以前）就已经有了以算法手段做出的非常非常惹人注目的图像，所以我不确定这又是什么鬼。

Even so, I don't doubt that you can invent AI that can figure out how to make something that has a lot if not all the characteristics of what we call "art," even the really brainy stuff. There will be a Turing test where you won't be able to tell what is made by a human and what is made by a computer intelligence, no doubt about it.
即使如此，我也不怀疑你能发明出（这样的）AI，它可以弄清如何制作具有许多即便不是全部我们都能称之为“艺术”特征的东西，甚至（包括）真正聪明的东西。还会进行一次图灵测试 （译注：科学界对此有争议），（届时）你将不能说出（分清）什么是由人类制造的、什么是由计算机智能制造的，毋庸置疑。

But really what "art" as a category means to us is an invention of the Romantic period in Europe. And what it tends to mean very specifically is "proof of human creative genius," which took on extra cult-like status in response to industrialization, as people tried to find ways to feel like they were holding onto their humanity in a fast-technologizing world. So, as humans have invented new tools of image-making and form-making—photography being the classic example—what tends to happen is that what we call "art" mutates to find a new way to convey "human creative genius."
但实际上于我们来说作为一个分类手段的“艺术”是欧洲浪漫主义时期的一项发明。它往往非常明确所指的是“人类创造性天赋的证据”，这另外呈现的狂热状态是对工业化的回应，就像人们设法在迅猛科技化的世界保持本性。因此，人类已经发明的图像制作和构图摄影的新工具成为经典的范例时 （译注：意为“过时”）——往往要发生的是（那些）我们称之为“艺术”的东西会突变为找寻（下）一种新的方式来表达“人类创造性天赋”。

So, it's like, "OK, the photographers do portraits; painting is about exploring color and form and expression now," which is what happened about when photography became mainstream at the end of the 1800s. And then, what also happens is that, some other artists figure out how to use the tool to convey the new standard of what "art" is, and you get something like "art photography." And that tends to be how it goes.
因此，它就像，“好了，摄影师是给人照相的，（而）现在的绘画是有关探究色彩、形式和表现的”，这在这十九世纪末摄影成为主流的时候就发生了。之后，又发生的是，另一些艺术家掌握了怎样使用此工具来表达何为“艺术”的新标准，接着你就知道了某种（名为）“艺术摄影”的东西。而且往往就是怎样。

Maybe the smarter and more creative our computers get, the harder it gets for artists to find new strategies to symbolize "human creativity." Maybe the idea of celebrating exceptional "human creativity", in fact, is dated. But I'm pretty sure that's what the cult of "art" and the cult of the "artist" means, even today, so in that sense, Deep Dream just represents a modest new displacement, or challenge that artists holding onto that tradition have to tangle with, that's all.
也许我们的计算机越聪明越具创造性，对于艺术家来说就越难想出新的策略来表征“人类的创造力”。也许赞美特别的”人类创造力”的想法，事实上，就是过时的。但我敢肯定这就是（对）“艺术”的狂热崇拜和（对）“艺术家”手法的迷信，即使在今天，从这个角度说，Deep Dream 仅仅是一个温和的新改变，或者对坚持传统的艺术家来说不得不直面的挑战，仅此而已 （也可以翻译为“终于说完了”，我翻译得快爆炸了）。



##3.1 第一节

Google Research Blog
《Google研究网志》其一

Inceptionism: Going Deeper into Neural Networks
Inceptionism：更加深入地探究神经网络

http://googleresearch.blogspot.com/2015/06/inceptionism-going-deeper-into-neural.html

Posted: Wednesday, June 17, 2015
发文日期：周三，2015-6-17

Posted by Alexander Mordvintsev, Software Engineer, Christopher Olah, Software Engineering Intern and Mike Tyka, Software Engineer
作者：
    Alexander Mordvintsev，软件工程师；
    Christopher Olah，软件工程师实习生；
    Mike Tyka，软件工程师

Update - 13/07/2015
更新日期：2015-7-13

Images in this blog post are licensed by Google Inc. under aCreative Commons Attribution 4.0 International License. However, images based on places byMIT Computer Science and AI Laboratory require additional permissions from MIT for use.
本博文中的图像为 Google 公司所有，采用知识共享署名4.0版本国际许可协议授权。不过，基于MIT计算机科学与人工智能实验室（做出的）的图像还需要另外向MIT申请使用许可。

Artificial Neural Networks have spurred remarkable recent progress in image classification and speech recognition. But even though these are very useful tools based on well-known mathematical methods, we actually understand surprisingly little of why certain models work and others don’t. So let’s take a look at some simple techniques for peeking inside these networks.
人工神经网络在 图像分类和 语音识别方面已经取得了显著的最新进展。但即使这些都是非常有用的基于知名数学方法的工具，我们实际上还是对为何某些模型能起作用而其他不能知之甚少。所以，让我们以简单的技术来窥探下这些网络里面（的东西）。

We train an artificial neural network by showing it millions of training examples and gradually adjusting the network parameters until it gives the classifications we want. The network typically consists of 10-30 stacked layers of artificial neurons. Each image is fed into the input layer, which then talks to the next layer, until eventually the “output” layer is reached. The network’s “answer” comes from this final output layer.
我们是这样训练一个人工神经网络的：向其展示数以百万计的训练样本并 逐步调整网络参数，直到它能给出我们想要的分类结果。该网络通常由 10~30 层堆叠的人工神经元组成。每幅图像都由输入层喂入 （送入），然后递入下一层，直到最后抵达“输出”层。网络的“答案”就来自于这个最终的输出层。

One of the challenges of neural networks is understanding what exactly goes on at each layer. We know that after training, each layer progressively extracts higher and higher-level features of the image, until the final layer essentially makes a decision on what the image shows. For example, the first layer maybe looks for edges or corners. Intermediate layers interpret the basic features to look for overall shapes or components, like a door or a leaf. The final few layers assemble those into complete interpretations—these neurons activate in response to very complex things such as entire buildings or trees.
神经网络的一个挑战是要理解在每一层到底都发生了什么事。我们知道在（经过）训练之后，每一层会逐步提取越来越高级的图像特征，直到由最后一层做出显示为何的决定（本质上）。例如，第一层也许在寻找边缘或拐角。中间层分析基本的特征来寻找整体形状 （即大体轮廓）或其组成部分，比如一扇门或一片叶子。最后几层将那些（要素）组合成完整的解释——（如此，）这些神经元对非常复杂的东西诸如整个建筑物或树木就有了反应。

One way to visualize what goes on is to turn the network upside down and ask it to enhance an input image in such a way as to elicit a particular interpretation. Say you want to know what sort of image would result in “Banana.” Start with an image full of random noise, then gradually tweak the image towards what the neural net considers a banana (see related work in [1], [2], [3], [4]). By itself, that doesn’t work very well, but it does if we impose a prior constraint that the image should have similar statistics to natural images, such as neighboring pixels needing to be correlated.
视觉化此过程的一种方法是将网络上下颠倒，并且要求它采用这种方法以引出一个特定的解释来增强输入的图像。假设 （比如说）你想知道什么样的图像会导致“香蕉”这样的结果出现。从一幅充满随机噪点的图像着手，然后逐渐调整图像朝着神经网络认定其为香蕉的方向变化（见相关工作： [1], [2], [3], [4]）。本来效果并不怎么好，但如果我们施加先验约束就不一样了，即图像应该有类似的对自然图像统计的信息，比如邻近像素需要相互关联。


So here’s one surprise: neural networks that were trained to discriminate between different kinds of images have quite a bit of the information needed to generate images too. Check out some more examples across different classes:
因而这里有一个惊喜：被训练用以区分不同类型图像的神经网络也有相当多的信息需要去生成图像。查看一下（下面）更多不同类别的例子：


Why is this important? Well, we train networks by simply showing them many examples of what we want them to learn, hoping they extract the essence of the matter at hand (e.g., a fork needs a handle and 2-4 tines), and learn to ignore what doesn’t matter (a fork can be any shape, size, color or orientation). But how do you check that the network has correctly learned the right features? It can help to visualize the network’s representation of a fork.
为什么这个很重要？呃，我们通过简单向它们展示很多我们想让其学习的范例来训练网络，希望他们提取所关注事物的要素（例如，一个分叉需要一个柄和2~4个尖齿），并学会忽略不相关的东西（分叉可以是任意形状、大小、颜色或方向）。但是如何检查网络（是不是）已经学到了对的特征呢？它可以帮助视觉化网络一个分叉的表达。

Indeed, in some cases, this reveals that the neural net isn’t quite looking for the thing we thought it was. For example, here’s what one neural net we designed thought dumbbells looked like:
的确，在某些情况下，这揭示了神经网络并不是在寻找我们以为它在（寻找的）东西。例如，下面这个我们设计去思考哑铃的神经网络看起来像：


There are dumbbells in there alright, but it seems no picture of a dumbbell is complete without a muscular weightlifter there to lift them. In this case, the network failed to completely distill the essence of a dumbbell. Maybe it’s never been shown a dumbbell without an arm holding it. Visualization can help us correct these kinds of training mishaps.
在那儿是有哑铃，但似乎没有一个图像中的哑铃是完整的，那儿并没有一个肌肉发达的举重运动员要来举起它们。在这种情况下，网络未能完全提取到哑铃的要素。也许它还从来没被展示过一个无手臂握着的哑铃（图片）。视觉化能帮助我们纠正这些训练事故。

Instead of exactly prescribing which feature we want the network to amplify, we can also let the network make that decision. In this case we simply feed the network an arbitrary image or photo and let the network analyze the picture. We then pick a layer and ask the network to enhance whatever it detected. Each layer of the network deals with features at a different level of abstraction, so the complexity of features we generate depends on which layer we choose to enhance. For example, lower layers tend to produce strokes or simple ornament-like patterns, because those layers are sensitive to basic features such as edges and their orientations.
替代确切地规定我们想要网络去放大的特征，我们也可以让网络（自己）做决定。在这种情况下，我们简单地给网络喂入任意的一幅图像或照片，并让网络分析。然后我们挑选出一层，要求网络增强任何它能检测到的东西。网络的各层（神经元）会处理图片中不同抽象程度的特征，所以我们生成的特征其复杂性取决于我们选择去增强哪一层。例如，较低的层会倾向于产生笔触或者简单的装饰性图案，因为那些层对基本的特征比如边缘和它们的方向很敏感。



If we choose higher-level layers, which identify more sophisticated features in images, complex features or even whole objects tend to emerge. Again, we just start with an existing image and give it to our neural net. We ask the network: “Whatever you see there, I want more of it!” This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird. This in turn will make the network recognize the bird even more strongly on the next pass and so forth, until a highly detailed bird appears, seemingly out of nowhere.
如果我们选择更高级别、可以识别图像中更复杂特征的层，（那么）复杂的特征抑或整个对象往往就能显现出来。还有，我们只是以现有的图像开始，将其递入我们的神经网络中。我们要求网络：“不管你看到的是什么，我都想要更多！”这样就创建了一个反馈循环：如果云朵看着有点像一只鸟，（那么）神经网络将会把它变得更像鸟。这反过来将使网络在下一步以及更多步识认鸟时更加强烈，直到出现一个高度精细（具体）的鸟，（而它）似乎不知是从哪儿冒出来的。


The results are intriguing—even a relatively simple neural network can be used to over-interpret an image, just like as children we enjoyed watching clouds and interpreting the random shapes. This network was trained mostly on images of animals, so naturally it tends to interpret shapes as animals. But because the data is stored at such a high abstraction, the results are an interesting remix of these learned features.
（研究）结果很是有趣——即使一个相对简单的神经网络也可以被用于过分解读图像，就像孩童时代我们喜欢看云朵、解读（出现的）随机形状一样。这个网络（之前）主要是以动物的图像被训练的 （另译：由于之前主要是以动物的图像来训练这个神经网络的），因此自然它就倾向于把（一些）形状解释成动物。但由于数据被存储的如此高度抽象，结果就出现了这些习得特征的一个有趣的杂糅。


Of course, we can do more than cloud watching with this technique. We can apply it to any kind of image. The results vary quite a bit with the kind of image, because the features that are entered bias the network towards certain interpretations. For example, horizon lines tend to get filled with towers and pagodas. Rocks and trees turn into buildings. Birds and insects appear in images of leaves.
当然，我们可以做的远不止运用了这种技术的云朵。我们还可以把它应用（推广）到任何种类的图像。（而）一种图像的（处理）结果大异，是因为输入的（众多）特征造成了（神经）网络的偏差，使其朝着某些（方向）解析。例如，地平线往往会充斥着塔楼和宝塔 （佛塔），岩石和树木变成了建筑物，鸟类和昆虫出现在叶子的图像中 （即 AI 对叶子和虫鸟有点傻傻分不清）。


This technique gives us a qualitative sense of the level of abstraction that a particular layer has achieved in its understanding of images. We call this technique “Inceptionism” in reference to the neural net architecture used. See our Inceptionism gallery for more pairs of images and their processed results, plus some cool video animations.
这种技术给我们提供了一种对抽象层次的定性感受，即特定的某层已经取得了对图像（怎样）的理解。根据运用的 神经网络体系结构，我们称这种技术为“Inceptionism”。想要更多对图像及其处理结果的图片，参见我们的 Inceptionism 相册，里面还有一些很酷的视频动画。

We must go deeper: Iterations
我们须得更深入些：迭代

If we apply the algorithm iteratively on its own outputs and apply some zooming after each iteration, we get an endless stream of new impressions, exploring the set of things the network knows about. We can even start this process from a random-noise image, so that the result becomes purely the result of the neural network, as seen in the following images:
如果我们把此算法反复应用在其自身的输出上，并在每次迭代后应用一些缩放，我们就能得到了无尽的新效果（数据）流，探索网络了解（知晓）的一众事物。我们甚至可以从一幅随机噪点 （噪声）图像开始这个过程，以便（经处理后的）结果完全变成神经网络的结果，如下面的图像所示：


The techniques presented here help us understand and visualize how neural networks are able to carry out difficult classification tasks, improve network architecture, and check what the network has learned during training. It also makes us wonder whether neural networks could become a tool for artists—a new way to remix visual concepts—or perhaps even shed a little light on the roots of the creative process in general.
这里介绍的技术（可以）帮助我们理解以及视觉化神经网络是如何执行困难的分类任务的，改进网络体系结构，并检查在训练期间网络都学到了什么。它还让我们好奇神经网络是否可以成为艺术家的工具——一种再混合视觉概念的新方法——或许即使通常（只有）一点点光透到创意过程的根基上。

Labels: Computer Vision, Image Classification, Neural Networks
标签：计算机视觉、图像分类、神经网络

---

##3.2 第二节

Google Research Blog
《Google研究网志》其二

DeepDream - a code example for visualizing Neural Networks
Deep Dream，可视化神经网络的代码范例

http://googleresearch.blogspot.com/2015/07/deepdream-code-example-for-visualizing.html

Posted: Wednesday, July 01, 2015
发文时间：周三，2015-7-01

Posted by Alexander Mordvintsev, Software Engineer, Christopher Olah, Software Engineering Intern and Mike Tyka, Software Engineer
作者：
    Alexander Mordvintsev，软件工程师；
    Christopher Olah，软件工程师实习生；
    Mike Tyka，软件工程师

Two weeks ago we blogged about a visualization tool designed to help us understand how neural networks work and what each layer has learned. In addition to gaining some insight on how these networks carry out classification tasks, we found that this process also generated some beautiful art.
两周前我们发表了一篇博文，其中阐述了一个可视化（视觉）工具被设计用来帮助我们理解神经网络的工作原理以及每层学到了什么东西。除为了获悉这些（神经）网络是如何进行分类任务 （作业）的，我们还发现这个过程也可以生成一些漂亮的艺术作品。


We have seen a lot of interest and received some great questions, from programmers and artists alike, about the details of how these visualizations are made. We have decided to open source the code we used to generate these images in an IPython notebook, so now you can make neural network inspired images yourself!
我们看到了（大众对此的）极大兴趣，也收到了一些好的问询，来自程序员和艺术家们的（反馈）差不多，都问到了这些视觉化怎样实现的细节。我们已经决定开放在 IPython notebook 中用于生成这些图像的源代码，因此现在你可以自己做些神经网络灵感的图像了！

The code is based on Caffe and uses available open source packages, and is designed to have as few dependencies as possible. To get started, you will need the following (full details in the notebook):
此代码基于 Caffe 并使用了可获取的开源（软件）包，也设计尽可能地少用依赖。要开始使用它，你还需要下列的东西（手册里含有所有的细节）：

    ·NumPy,SciPy, PIL, IPython, or a scientific python distribution such asAnaconda orCanopy.
    ·Caffe deep learning framework(Installation instructions)
    #-------------------我---是---分---割---线-------------------
    ·NumPy、SciPy、PIL、IPython、像Anaconda或Canopy这样的用于科学研究的python分发包；
    ·Caffe深度学习框架（安装指导）

Once you’re set up, you can supply an image and choose which layers in the network to enhance, how many iterations to apply and how far to zoom in. Alternatively, different pre-trained networks can be plugged in.
一旦配置好了，你就可以输入一幅图像，选择要增强网络中的哪些层、迭代多少次、把画面拉到多近 （即放大）。而且，你还可以选择（经历）不同预先训练的网络。

It'll be interesting to see what imagery people are able to generate. If you post images to Google+, Facebook, or Twitter, be sure to tag them with #deepdream so other researchers can check them out too.
看看人们能做出何种图像会很有意思，如果你要在 Google+、非死不可或推特上发布图片，记得加上 #deepdream 标签。如此，其他人就也能切克闹一下。

Labels: Computer Vision, Neural Networks, open source
标签：计算机视觉、神经网络、开源



##4.1 英国电讯报图片新闻

###4.1.1 正文：

Google Deep Dream: 19 of the best images from mesmerising photo software
Google Deep Dream：从这个梦幻般的图像软件中精选出的 19 幅最佳图片

http://www.telegraph.co.uk/technology/google/11730050/deep-dream-best-images.html


###4.1.2 评论（Comments）：

Inigo Montoya. • 8 days ago
What utter rubbish. Pointless.
真垃圾，毫无意义。

Dave Casey • 14 days ago
New word: Hallucigeneric!
新词（产生了）：迷幻类的！

D. Briggs • 16 days ago
check out Facebook Group Google DeepDream Imagery. I think there are lots of great ones there.
去看看非死不可群组（团）——Google DeepDream 图像，那里有很多好图。

Sirkoes • 16 days ago
Someone at Google read too much Lovecraft books.
Google 里某些人一定是看了洛夫克拉夫特太多的小说。

Dick_Turpin • 18 days ago  +1
Someone feed it some HR Giger art, and tell us if it turns your computer psychotic. ;-)
希望有人能给它喂入汉斯·鲁道夫·吉格尔 （瑞士知名超现实主义艺术家，曾设计电影《异形》中的外星生物，因而赢得奥斯卡金像奖的最佳视觉效果奖，已经离世。）的作品，然后告诉大家你的电脑有没有得精神病。-_-||

Kaloyan • 20 days ago
Want to do some of my own. Found this one - http://deepdreamgenerator.com/ but it is a bit slow. Is there anything faster?
想做点自己的图，请访问这个网址—— http://example.com，不过有点慢，谁有更快的？

Mike Hunt-Hurtz • 21 days ago  +2
#19 very hallucinate such acid
第 19 幅图像太迷幻了

traveloffice • 24 days ago  +1
what a pure crap. it's a piece of crap software witch adds random eyes to every image.
完全是垃圾。这是由垃圾软件随机地加一些眼睛到每幅图里。

    Dave Casey traveloffice • 14 days ago
    What a pure crap it's a travel that adds office to every travel office that adds travel office travel office traveloffice trailoff trailoff
    你也是个垃圾，是个垃圾，个垃圾，垃圾，圾...... （译注：模拟回音，此人在攻击层主的 ID）

    keka traveloffice • 23 days ago  +2
    someone is afraid of eyes
    有些人确实害怕那些眼睛

         HalfMooner keka • 9 days ago
        The guilty always are.
        心虚的表现。

Beadybonce • 25 days ago
So they stuck fractals on the photos?
图上的东西尽是支离破碎 （乱糟糟）的？

    Daiyu Hurst Beadybonce • 21 days ago
    The world is a fractal.
    世界本来就是支离破碎的。

Mr Douglas Dixon • 25 days ago
Andy Murray's has not changed anything? That's how he looks.
安迪·穆雷的那幅没怎么变化呀，他就长那样儿。

Roger • 25 days ago
I tried it on some of my prog rock album covers - nothing happened!
我拿自己的一些前卫摇滚唱片封面试了试，啥都没发生！

Harry • 25 days ago  +1
If these are the best, I'd hate to see the worst.
Complete waste of time.
若然这些就是最好的，我会厌恶看到那些最差的。

curious • a month ago
You can make your own so easily here: dreamingwith.us
在这里，你可以很容易地自己做些图 --> example.com

    This comment was deleted.
    该评论已被删除

         curious Guest • 25 days ago
        Why is that?
        什么情况？

Pressed Rat and Warthog • a month ago  +2
The '60s called and wants its software back.
60后的（老年人给 Google）打了电话，要求他们召回软件。

Gandalf-the-White • a month ago
Deep Dream likes using dogs. ?Exactly creating what looks like a dog's dinner most of the time..... use free Gimp easier better.
Deep Dream 很喜欢弄些狗狗？尽是些狗头大餐......劝大家还是使用免费的 GIMP （跨平台的开源版 Photoshop）软件，它上手更容易、效果也更好。

elhongo • a month ago
I have the Google Deep Dream up and running. I can process any image you need. https://www.fiverr.com/elhongo...
我配置好了 Google Deep Dream 并架设了服务器，我可以为你处理任何图像。请访问： https://www.example.com/nofree

curious • a month ago
I made a site that will let you create your own Check it out- http://dreamingwith.us/ this one processes it much faster :-)
我做了个网站，它可以让你创作自己的图片，来瞧瞧 --> http://example.crashed，它（比其他人）处理的更快些。

    Діма Крищук curious • a month ago
    but your site doesn work(
    然而，你的网站（压根就）上不去。

         curious Діма Крищук • a month ago
        Aaah fixed :-)
        啊哈，修好了。 （译注：目前还是上不去）

Angus Gordon-Farleigh • a month ago
Neural network: Object recognition.
神经网络：物体（对象）识别

Joe_Pubic • a month ago  +1
and the point of this programme is?
这个程序的要点 （原理）是什么？

    itzman Joe_Pubic • 25 days ago  +2
    Well originally it was to train software to see patterns: they then played around with the gain making it see patterns that weren't really there.
    The similarity to LSD experiences is so marked that it probably tells us rather a lot about the effect of hallucinogenic drugs, and possible a bit about some mental illnesses.
    The point was to create software, the unexpected result was what happened when the software was given the software equivalent of LSD.
    Its extremely interesting to people who are studying perception.
    最初是为了训练软件去识别图形，然后他们用取得的成果乱搞，让软件识别本不在那儿的图形。
    这非常类似于摇头丸的体验，它可能（间接）告诉了我们关于迷幻药反应的很多东西，可能还些许涉及精神疾病的方面。
    其目的是开发出一款软件，而其意料之外的结果都由软件被给定的等量摇头丸所致。
    这对那些正在学习觉察 （译注：机器学习的范畴）理论的人来说很有意思。



##5.1 第一节 要点整理

###5.1.1 Deep Dream 项目地址：

https://github.com/google/deepdream

###5.1.2 谷歌 Deep Dream 看到的照片为什么有那么多狗头？

解释：因为 ImageNet1000 的数据集里面几百类的狗、几十类的鸟，统共超过一半都是动物，所以 Deep Dream 的世界就是个动物世界，或者说是个狗窝......
正经版回答：这个是因为数据集采样偏差的缘故，ImageNet 中类别的频率并不代表实际生活中我们见到的这些类别的频率，所以 ImageNet 的模型不能被直接用在实际产品当中。

译者注：这个问题的答案采自 Caffe的开发者（创造者）——贾扬清的知乎回答，鄙人可不敢乱说。贾扬清毕业于加州大学伯克利分校，现就职于 Google 公司。

###5.1.3 ImageNet 简介：

ImageNet 是一个计算机视觉系统识别项目， 是目前世界上图像识别最大的数据库。由美国哈佛的计算机科学家模拟人类的识别系统建立，能够从图片识别物体。 （摘自百度百科）

###5.1.4 Caffe 简介：

Caffe（卷积神经网络框架），全称为 Convolutional Architecture for Fast Feature Embedding，是一个计算 CNN 相关算法的框架。
Caffe 是一个清晰、可读性高、快速的深度学习框架，官网 http://caffe.berkeleyvision.org/。 （摘自百度百科）

###5.1.5 IPython notebook：

IPython 是一个加强版的交互式 Python Shell；在 2011 年，由 Brian Granger 领导的 IPython 团队开始开发一种基于 Web 技术的交互式计算文档格式，即 IPython Notebook。
IPython Notebook 使用浏览器作为界面，向后台的 IPython 服务器发送请求，并显示结果。Notebook 在交互上使用了 C/S 结构，它通过 Tornado 建立一个 shell 服务器，并使用浏览器作为客户端。另外 notebook 页面都被保存为 .ipynb 的类 JSON 文件格式，这种文件格式也是 Notebook 最吸引人的地方。
IPython notebook 目前已经成为用 Python 做教学、计算、科研的一个重要工具。 （摘自多篇网文）

---

##5.2 第二节 Reddit 社区问答 （仅摘取翻译，不过原主题非常活跃，评论也很多）

https://www.reddit.com/r/deepdream/comments/3cawxb/what_are_deepdream_images_how_do_i_make_my_own/

###5.2.1 How do I make my own?
如何才能自己动手做图？

Without programming experience:
无编程经验的情况下（有编程经验的已略去，想自己配置软件的可以移步上面给出的主题帖链接）：

（下面是可用的免费网站网址，提交图片后请等待服务端处理，需时各异）

1. http://dreamscopeapp.com

2. http://deepdream.in

3. http://deepdreamer.io

4. http://psychic-vr-lab.com/deepdream/

5. http://deepdream.akkez.ru

6. http://deepdreamit.com

7. http://deepdreamgenerator.com

8. http://deepdream.pictures/static/#/

9. Check out the subreddit where people fulfill your requests for you! just give them the image. /r/deepdreamrequests.（即请人代劳，类似于贴吧或论坛的“求P图”）

###5.2.2 Can this be done on audio? video?
音频和视频是否也能做？

Yes. To make a video you can run the code on each individual frame of the video then stitch them together afterwards. But there are more efficient ways discussed in this thread. The best resource for learning about this is here: https://github.com/graphific/DeepDreamVideo
可以。要想做视频，你可以在视频的每一单帧上运行这段代码 （软件），然后再合并到一起 （时间轴）。不过， 在这个帖子的讨论中有些更具效率的方法。了解此事的最佳来源是这里： https://github.com/graphific/DeepDreamVideo

If you wish to make one of those zoom-into-an-image-really-far gifs like this one then you should follow the guide here: (TODO: guide link)
如果你希望做些与 此类似的 gif 动态图片，那么你应该照着这里的指导做（备忘：教程链接）：

（其实现在还没有放出链接，所以才加的“TODO”字样，也可能忘了修改了。所以，我从别的段落中采集并补上这个链接，如下：）
https://github.com/samim23/DeepDreamAnim

To perform this on audio you need to really know what you are doing. Audio works better with RNNs than CNNs. You will need to create a large corpus of simple music to train your RNN on.
要想在音频上玩耍，你得真正地了解自己在做什么。音频在多层反馈 （递归）神经网络（Recurrent Neural Network）上的表现比在卷积神经网络（Convolutional Neural Networks）更好些，你还要创建大量的样本 （示例）音乐用来训练你的反馈神经网络。




变成狗头大餐的意大利面


肯德基外卖全家桶






二次元图片对比


对同一图片设置不同参数做出的不同风格图片


拳王


梵高的名作《星空》




壮丽的景色