黑人竟被认成大猩猩？图像识别靠谱吗？,创业美国微信公众号文章

在人工智能和机器学习领域中，谷歌堪称领军企业。但早前，其具有图像识别功能的产品Google Photos却误把两名黑人标注为“大猩猩”。随后谷歌迅速对此道歉，并表示将调整算法以修复这一问题。而Google Photos并不是第一个犯这类错误的图像识别工具。雅虎旗下的照片分享网站Flickr也曾经把用户上传的照片标记错误，错把纳粹集中营标记成“运动”。

基于机器学习技术领域的图像识别到底靠谱吗？

《创业美国》第5季第2期所采访的公司Clarifai，正是一家致力于利用深度学习算法来进行视觉搜索的图像识别公司。

Clarifai创始人Matthew Zeiler，纽约大学计算机科学PHD，是“深度学习”开山鼻祖Geoffrey Hinton的得意门生。为了创业，他几乎拒绝了整个硅谷。

2016年，Clarifai刚刚收获了价值 3000 万美元的 B 轮融资。作为图像识别领域为数不多的独立玩家，Clarifai在成立四年多的时间里率先将图形识别从静态图片带入了接近实时的级别。

究竟是什么样的黑科技，让Clarifai可以超越图像识别界的大牛谷歌，成为世界上唯一一家可以识别视频图像的公司？

－《创业美国》第5季第2集完整视频－

★ 本期节目中英访谈纪实★

创业美国陈一佳（以下简称C）

Clarifai创始人Matthew Zeiler （以下简称M）

图形识别的核心技术

C：What kind of technologies behind it

电脑识别图像的科技原理是什么

M：So we have artificial intelligence technology to understand images and video automatically. This just takes in the pixels of the image.It doesn't need to know anything else and it can predicts things like dog, tree, mountain. As well as descriptive words like happiness or scenic automatically from your data.

我们采用的是人工智能技术，这项技术能够自动让电脑识别图片和视频，只需要根据画面上的像素做判断，不需要别的就可以区分并识别出狗狗、树木和山峦也可以识别一些描述性的词，比如“快乐”或者“景色秀丽”这样的词汇，只要获得相应数据就可以自动识别。

C：What does that really mean?Do you think now AI is smarter than human being?

这背后的意义是什么?你是否认为人工智能要比人更聪明？

M：It depends on how you define intelligence and smartness of humans, because in certain applications machines can definitely outperform humans. So AlphoGo is one that they have built algorithms that can surpass the best humans on the planet And with image recognition. We're seeing the same thing, so we have algorithms that can surpass accuracy of even doctors.

这取决于你看你如何定义什么是聪明了。在某些领域内机器是一定超越人类的。比如阿尔法围棋，一旦建立了算法这个人工智能机器就可以打败全世界最厉害的围棋手。至于我们所在的图像识别领域，我们也在看到一样的趋势。我们在图像识别方面的正确率甚至超越医生。

M：So specialists trained for decade in school that are experts practicing every single day the reliability and the consistency of a machine can be much higher and it can also learn from much more data than a doctor is exposed to.

而这些医生可是在医学院里有过数十年学习，每天经过严格临床训练的专家。但是机器相比这些人类医生具有更高的一致性和可信赖度同时他们比医生获得的数据还有多得多。

M：For example, my dad is actually a doctor in the small farm town that I grew up in. And so the number of patients that he could possibly see in his lifetime is so many orders of magnitude smaller than how much you can feed through a machine. And so it has so much more opportunity to be smarter than a human.

比如我父亲就是一个医生。在我从小长大的那个一个农场小镇里工作，毕生看到的病人数量和机器能够接触到的数据比根本就不是一个数量级。所以相比人类，机器有更多变聪明的机会。

C：I remember Google system labeled african american couple as gorillas, so are there any similar limitations of Clarifai? Does Clarifai make any mistakes?

我记得谷歌的图像识别系统曾经把一对黑人情侣标识成了大猩猩,Clarifai有没有类似的局限，或者说Clarifai在识别中是否也会犯错？

M：I mean of course Clarifai makes some mistakes. We haven't had any kind of PR disaster like that and we have a blacklist of concepts that we would view as being offensive.

这是自然的，Clarifai当然会犯一些错误。不过目前为止，还没有发生过谷歌那样严重的公关性灾难。而且我们有一个黑名单，会把这些具有攻击性的概念囊括其中。

C：How accurate is Clarifai now?

Clarifai目前的准确率怎么样

M：Accuracy is kind of in the eye of the beholder, because it depends on what data you actually care about. We have coming in our platform is an ability to evaluate the accuracy on your specific data which is really important for all of our customers.

准确率这个东西是因人而异的，因为它的关键在于用户最看重的是哪方面的数据。在我们的平台上，用户最看中哪部分的数据，我们会在那一部分数据里搭建正确率评估体系，这一点对于我们的用户来说是至关重要的。

C：How fast it can be?

识别的速度有多快？

M：In a fraction of a second , just take a picture and find the most visually similar stuff.

几秒钟的时间，你只要拍一张照片，系统就能快速识别与图片物品类似的图像。

C：How about other competitors, for example Google system, how fast can they do?

那你们的竞争对手呢，比如谷歌他们的图像和视频识别系统能做到多快？

M：They actually don't support video today.

他们其实还不支持视频识别。

C： are the only guy?

你们是行业里唯一能够识别视频的公司吗？

M：Yes.

没错。

图像识别的应用领域

C：How important is image recognition?

图像识别为什么重要？

M：If you want to discover content, there is no way you could flip through the whole internet, That’s just impossible. But a machine could, that applies anywhere you have an image or if video in every possible industry. Where there's an image video, image recognition can understand it and extract value from it.

如果你想要在网络当中寻找与一个主题有关的全部内容，你不可能翻遍整个网络而毫无遗漏，因为这是不可能做到的。但是机器可以。对于任何有图片和视频的用户，对于任何产品，只要他们有图片和视频方面的识别需求，就能够明白我们的价值有多大。

C：What kind of customers do you have?

你们有多少客户

M：One is organization and one is moderation. So organization is this discovery experience where all the content from a company and their end users. Creating new content can be organized so that other users can come and search and find the content they're looking for.

我们的第一种客户用我们进行搜索整合，第二种客户用我们进行管理和控制。第一种客户获得的是一种发现的体验，也就是说网络上有关这个公司的，以及来自终端用户的一些新的内容都可以通过我们被搜索和整合。这样的话，他们的用户就可以在网络上搜索他们需要的相关内容。

M：For example, where the person creating the content is a professional photographer and the buyer might be a head agency or something like that. And they want some good content, and so both of them benefit. The uploader can use our technology to tag the content much faster than they can manually tag it. Which they had to do in the past and the buyer now gets a better experience because every single image is tagged in the same way. Maybe you can search and find wedding content easier or a travel site and trying to find your next vacation destination.

举个例子，创作内容的人。比如一个职业摄影师，和购买内容的人。可能是某个公司的志愿，或者类似的角色。他们想要购买一些好的内容（照片），这两个角色都可以通过我们获益。摄影师作为图片的上传者可以利用我们的技术为自己的作品自动贴标签，这比他们手动一个一个贴要快速得多。但是过去每个摄影师都不得不这样做。对于想要购买这些相片的人来说呢，购买的体验会更好，因为每一张同属的相片都以同样的方式做了标注。比如你想要搜索与婚礼有关的图片，或者在旅行网站上，搜索你的下一个度假地，你都可以快速发现和找到你想要的内容。

M：Now the second bucket, I was mentioning is moderation and we have a lot of customers using us to filter out content that they don't want on their site. Either internally or like a marketplace or a public part of their site. That big brands themselves like Unilever are using us and so they want to have their marketing team find throughout social media good examples of pictures that are being created that convey their brand.

我刚才说到的第二类客户是管理型客户。很多客户通过我们在网络当中筛查出他们不想要的信息，无论是公司内部网络、还是一个电商、或是网上的一个公开网页（都有类似需求）。还有不少大公司，比如联合利华也是我们的客户。他们想要让自己的市场部门找到社交网络上有关联合利华品牌的的一切图片。

M：And so the marketing team is able to use our new training technology which includes a user interface they don’t have to write a single line of code. And then once the model is trained it can go through all social media and find the pictures for them that they can use in their next ad campaign. And so some offensive things like nudity drugs weapons violence, we can recognize all these things and prevent them from being displayed on your site.

那么这个市场部门就可以利用我们的新技术，也就是一个用户界面。不需要书写任何代码，一旦训练模型建成，他们就可以在整个社交网络当中寻找他们需要的图片，并且用于他们下一次的广告推广当中。另外一些负面的东西，比如黄色、武器或者毒品等，我们也可以帮助筛选出来。这样他们的网站上就不会有这些内容出现了。

C：You are not alone in the field, do you think Clarifai can be better?

这个行业显然不是只有你们一家企业，相比竞争对手你们是否更优秀呢？

M：Yes. And a lot of our customers compare us to the other publicly available APIs. Some of those companies you mentioned don't have APIs. Apple is another example who has publicly said they're investing a lot in machine learning, but they don't have it. We're an API platform that people across all different industries can use to recognize their visual content.

没错，很多人会把我们和一些公开拥有API的公司相比较。但是还有很多大公司没有开发API。比如苹果，他们公开说过，在机器学习方面投入很大，但是却还没有推出API平台。我们是一个典型API平台，让各行各业的人，都可以识别他们行业当中需要的视觉信息。

M：And it's very simple, Send us an image, we tell you what's in it. Or send us a video we tell you the time series of what's in it. Google, Microsoft, IBM and Amazon, all four of them have APIs, and they are now our competitors.

方法也很简单，发给我们图片，我们就会告诉你图片上有哪些元素。或者发一个视频给我们，我们会告诉你不同时间段内显示了哪些信息。谷歌、微软、IBM以及亚马逊，这四家公司都有API，他们才是我们的竞争对手。

被硅谷争抢的最强毕业生

C：How did you get into AI and image recognition？

你是如何进入到人工智能尤其是图像识别这一领域的？

M：It happened when I was in undergrad at University of Toronto. So Jeff is kind of the pioneer of this field has been working on these neural nets since the 70s and 80s and he was an emeritus professor. It was the first time he’s ever taken an undergrad. So I got a taste in machine learning in undergraduate. I didn't know enough about machine learning in order to start a company, and that's what brought me here to New York to go to NYU to learn more about neural networks applied to images. So I did my Ph.D. at NYU for four years.

我当时在多伦多大学读本科，Jeff是这个行业的先锋。他从70年代80年代就开始研究神经网络，是大学里的名誉教授。后来他把我招进了自己的研究项目，那还是他第一次接收一个本科生进入项目。所以早在本科期间，我有机会了解到机器学习的皮毛，但是显然当时的知识储备不足以支撑我在这方面进行创业。所以我从加拿大来到纽约，就是为了可以在纽约大学继续学习去更深刻地了解神经网络在图像识别当中的应用，所以在纽约大学读了四年博士。

C：You interned at Google？

你曾经在谷歌实习过？

M：Yep! At Google Brain.

没错！“谷歌大脑”项目

C：Tell us about your experience there.

当时体验如何？

M：I spent the summer of 2012 and 2013 at Google Brain and I was hosted by Marc'Aurelio Ranzato who is now at Facebook AI. And then the second summer I was hosted by Jeff Dean who runs the Google brain team. One of my projects that actually made it out to the real world was a house number recognition system. That was really cool because after my internship, when I was using the Internet, in certain sites it would pop up. And I was like oh, I built that!

我2012年和2013年的暑假都在“谷歌大脑”实习。第一年跟随Marc'Aurelio Ranzato，他现在是Facebook人工智能方面的研究科学家。第二年夏天，我师从Jeff Dean，他负责运营整个“谷歌大脑”项目。当时我做的一个项目后来真的问世了，那是一套谷歌地图用的门牌号码识别系统。这个感觉很酷。后来我实习结束上网的时候，时不时会跳出来这个系统。我会想说，这就是我当年做的。

C：I heard you got so many offers after you graduated.

我听说你毕业以后拿到了很多工作Offer。

M：Couple weeks after I came back, I got a phone call, and it said on my call display Alan Eustace, and I have never talked to him before. He was a senior vice president at Google of their entire search and knowledge division. And I was like OK, I pick up and he said I talked to Jeff Dean, we really want you to join Google when you graduate. So we put together the largest new graduate offer we've ever given a student.

我实习结束回来几周后接到一个电话，手机上显示打电话的人是Alan Eustace。我从来没有跟他说过话，他是谷歌的副总裁，负责整个谷歌知识研发部门，于是我就接了电话。他说我和Jeff Dean聊过了，我们诚挚邀请你在毕业后加入谷歌，我们还为你开出了谷歌历史上给应届毕业生的最丰厚Offer。

C：How much money？

多少钱？

M：I can't say. I was like wow this is crazy. And I knew Microsoft, Facebook and Apple wanted to chat as well, so I did a low road trip and met everybody, I met Mark Zuckerberg in the process which is a really interesting.

我不能说。但是我当时被那个数字震惊了。但我知道微软、Facebook和苹果也都有意跟我聊一聊，所以我做了一次短暂的公路旅行，见了不少人。中途还见过扎克伯格，很有趣的经历。

C：How is he? what’s your first impression about him?

他怎么样？你对他的第一印象是什么？

M：Incredibly smart and humble as well And the way he speaks is really inspiring. It was a really cool opportunity to meet him.

极其聪明的一个人，也非常谦虚。他说话很激励人心，那是一次很棒的机会。

C：So he gave you an offer?

所以他给了你Facebook的工作offer？

M：So I got an offer from Facebook from Apple and Microsoft ended up doubling Google's offer. And then Google matched that, and put a 24 hour window on it. So I had this crazy period where I had a lower risk of joining one of the tech giant

我拿到了Facebook和苹果的offer，后来微软又在谷歌的条件上翻了倍，然后谷歌又加到了一样的条件，给我24个小时时间回复。那段时间很疯狂，一方面我可以加入这些独角兽企业，工作稳定没风险。

坚持理想一心创业

C：It must be hard to say no to these big guys.

对这些巨头说不，一定很难吧。

M：My parents were like take the money, but my wife and myself we knew what we had. What I realized was that the technology I had in New York was working better than what Google had internally. And Google was already years ahead of everybody else out there. And then I had these the early workings of Clarifai, working in my apartment and I submitted the first five neural networks that I had trained to this annual competition called ImageNet.

我的父母让我把钱收下好好工作。但是我妻子和我，我们知道自己手上有什么。我当时意识到我在纽约的技术，从内部上来说要胜过谷歌。但是谷歌相比别的公司，技术已经领先了好几年。当时我已经在自己的公寓里，搭建早期的Clarifai。我给一个叫做ImageNet的年度比赛提交了我训练的5个神经网络项目。

M：And the goal of the competition is to recognize a thousand different categories of objects within images. And three weeks later the results came out and clarify won the top five places in the competition. So it was a great way to kick off a company with the world's best image recognition. And so everybody at Clarifai realizes that they could go work at any of these companies, but we have a mission where we want to be the independent AI company. We don't want to get acquired, we want to build something big.

这个比赛是要你通过自己的技术识别出图片当中出现的上千个物品种类。三个星期过去答案揭晓，我们的识别正确率位于前五。作为图片识别当中的佼佼者，用这个成绩在这个领域创业是最好的方式了。在我们公司，所有人都知道他们可以去给任意一个科技巨头工作，但是我们有自己的任务。我们希望能够坚持做一家独立的人工智能公司，我们不想被收购，我们想要做一票大生意。

M：One of the coolest things I thought machine learning part was gonna be the coolest thing, but coming into the office and seeing 45 people now, all working our way towards our mission is incredible to see. So I'm very glad I decided on that.

我曾经想，机器学习技术已经是我们最酷的东西了。但是每天来到办公室，看到面前的45个人，都在朝着同一个目标努力，是很让人振奋的事情。所以我很高兴当出自己做了对的决定。

－《创业美国》第5季第2集完整视频－

★

访谈记录依然让你意犹未尽？

还有其他问题想问Clarifai创始人？

扫码关注创美并留言

下期《创业美国》为你答疑

★