English中文
将人工智能用起来
2019年6月18-21日
北京,中国

打造A.I.闭环 引领产业变革

此演讲使用中文 (This will be presented in Chinese)

温浩 (云从科技)
16:2017:00 Thursday, June 20, 2019

必要预备知识 (Prerequisite Knowledge)

对人工智能有初步的接触,希望了解更多的观众以及对人工智能企业定位感到困惑的创业者。

您将学到什么 (What you'll learn)

理解人工智能企业发展的脉络,了解未来中国人工智能发展的趋势。

描述 (Description)

云从是一家从中科院孵化出来的企业,我们的技术累积已有十多年,创始人周曦博士师从美国计算机视觉之父黄煦涛教授。除此之外,我们也得到了各级领导的重视,参与制定了国家标准和行业标准,承接了国家人工智能三大平台:发改委人工智能基础资源公共服务平台、高准确度人脸识别系统产业化应用平台,还有工信部核心SoC芯片应用平台。另外,我们在国内还设立了五大研发中心。

目前,我们的业务主要聚焦四个行业,银行、安防、机场和零售。 我们的愿景是成为国家级平台,以能力和资源为支撑点,凭借AI定义设备、定义场景做各个行业的解决方案,真正做到定义智慧生活。

前一段有一个新闻,宁波的行人闯红灯被抓拍,结果是公交车上的董明珠广告。这个事情告诉我们,AI技术在应用的时候存在很多问题,不能为了AI而做AI,它应该是有感知、有认知、有决策的一个闭环。或者你的技术够先进,能够形成一整套的有价值的方案。

云从要打造的就是技术闭环,具备感知、认知和决策能力。 人工智能是"头雁",人工智能解决的是问题的本身,怎样让能力提高,怎么去解决问题,让它引领技术发展是我们当前需要思考的主要问题。

当前,在技术落地方面,除了语音识别,人脸识别是应用最广的技术,很多人脸识别技术已经成为了人机交互的视觉入口。

其实在视觉识别中,除了人脸,还有人体识别,通过体态、衣着来识别人识别等,现在也开始有大规模的应用。

人脸加人体是对人识别的比较完善的方案。我们在中科院做了很多实验,对人脸做了各种角度、各种光源的分析,形成一个结构化的数据。

现在人体的识别率也超过了96.6%的商用标准。

举个例子,一个女孩在公园里面跑步,没有拍到人脸,但我们可以通过形体特征把她识别出来,这是一个跨摄像头和无需人脸的识别应用,所以我们叫它“跨镜追踪”技术。

公认的,人机交互的下一代交互方式就是“人脸+人体+语音+AI”,比如VR交互。那么云从在视觉识别外,还与联合实验室、中科院做了语音识别,错词率很低。同时我们在决策方面做了很多模型,比如双塔神经网络。

五官感知方面,我们通过感知技术做了统一大数据建模,通过机器人学习画像,得到策略推荐,再到执行、反馈。 比如"董明珠"闯红灯事件,就是感知系统不够完善,如果系统知道董明珠不可能出现在宁波,就完全可以避免这个问题。当然,这是一个比较极端的例子。

我们还做了行业产品和解决方案,在银行有50多个解决方案落地,通过感知技术做集成生物识别, 比如ATM机刷脸取款等,这是比较简单的应用。除此之外,银行对备付金预测也很关心,比如为某地建行的1000多个网点ATM机预测备付金,1个月超10亿的话就能节省上百万的利息,这是从感知到认知的决策闭环。

零售行业线下门店方面,从店外进来多少人,进来多少次、流连于什么产品间等,我们都有一套客户转化率去做识别和决策。 举个例子,门店门口装有人脸抓拍机,可以判断是VIP、会员、还是熟客,并根据性别年龄推送广告;到货架上,有感应技术获取画像并将其推荐到店员的终端上去,由他/她为顾客做相关的推荐;最后的支付环节,直接刷脸支付就可以了。

这才是现在AI技术可以帮助门店做的事情,而不是弄一个无人超市,对提高产量,毫无帮助。

最后我们认为,AI落地将从最开始的学术研究,体现技术先进性,一步一步走到行业验证,再行业实战,最后到成为行业平台和智能生态。

There was a news story recently. In Ningbo, a “pedestrian” was detected for crossing at a red light. But the truth is that the “pedestrian” was just an advertisement on a bus showing Mingzhu Dong, a famous entrepreneur in China. As this incident makes clear, AI technology has many problems when it’s applied in the real world. We can’t just use AI because we want to. AI should be a closed loop with perception, cognition, and decision making. Only then your technology is advanced enough to form a set of valuable solutions.

Hao Wen discusses how CloudWalk is building a closed-loop technology with the ability to perceive, recognize, and make decisions. Artificial intelligence is the “lead goose.” AI solves the problem itself. How to improve the ability, how to solve the problem, and how to let it lead the development of technology are the main problems we need to think about now.

At present, when talking about technology to applications, besides speech recognition, face recognition is the most widely used technology. And many face-recognition technologies have become the visual portal for human-computer interaction. In fact, in the visual recognition domain, in addition to face recognition, there is recognition of people by body pose and clothing, which now is also widely adopted. Recognition via the human face plus human body pose is a relatively better solution. Experimentation at the Chinese Academy of Sciences analyzed various aspects and light sources for the human face and then formed a set of structured data. Now the success rate of human body recognition is better than 96.6%, which is the standard for commercial use of this technology. For example, if a girl was running in a park and her face wasn’t caught by a camera, we can identify her through body features. This is a cross-camera and no-face-needed recognition application, so it’s called “cross-camera tracking” technology.

It is widely recognized that the next generation of human-computer interaction is “face + body + voice + AI,” such as VR interaction. Besides visual recognition, CloudWalk also conducted speech recognition research with the Joint Lab and the Chinese Academy of Sciences, whose error rate was very low. At the same time, many models in decision making have been built, such as the two-tower neural network.

In terms of five senses, CloudWalk has built a unified big data model through perceptual technology, which lets robots learn portraits, get strategic recommendations, and then implement and provide feedback. For example, the “Mingzhu Dong” red light crossing incident is due to a less perfect perception system. If the system knows that Mingzhu Dong is unlikely to appear in Ningbo, it can completely avoid this error. Of course, this is an extreme example.

We also developed industry products and solutions, implementing more than 50 use cases in the banking industry. For example, using perception technology, we built biometrical recognition, such as cash withdrawal in ATMs with face recognition. But this is a relatively simple application. In addition, banks are also very interested in forecasting the reserve fund. For example, a CCB branch needs to predict the reserve for more than 1,000 ATMs. If the forecasting can reduce RMB 1 billion reserve in one month, it will save millions of RMB in terms of interest. This is a good example of the closed loop of decision making from perception to decision.

In terms of offline stores in the retail industry—for questions like how many people entered a store, how many times they entered, what products they looked at—we have a set of customer conversion rate tools for identification and decision making. For example, stores are equipped with face-catch cameras, which can tell whether a caught face belongs to a VIP, a member, or a regular customer. And then stores’ system can push advertisements based on the customer’s age and gender. And on the shelf, sensor technology can catch a portrait and send it to a clerk’s terminal. He or she makes relevant recommendations for the customer. In the final payment step, showing a face would directly make the payment.

This is what AI technology now can do for offline stores, rather than building an unmanned supermarket, which does nothing for business improvement. Finally, CloudWalk believes that the AI-to-application process would from the very beginning start from academic research, showing the advancement of technology, to industry verification step by step, then to actual industry application, and finally to an industry platform and an intelligent ecology.

Photo of 温浩

温浩

云从科技

温浩,云从科技联合创始人。2003年获得中国科大电子科学与技术专业学士,并保送中国科大中科院量子信息重点实验室硕博连读,师从“量子调控”973首席科学家郭光灿院士,专攻量子通信器件和网络方向。2008年获得中国科大通信与信息系统博士学位,2014年加入中国科学院重庆绿色智能技术研究院。2015年和周曦博士共同创立云从科技。

Hao Wen is a cofounder, along with Xi Zhou, of CloudWalk Technology. Previously, he worked for the Chongqing Institute of Green Intelligent Technology of the Chinese Academy of Sciences. He holds a bachelor’s degree in electronic science and technology from China University of Science and Technology, where he was recommended for admission into the doctoral program at the China University of Science and Technology’s Key Laboratory of Quantum Information; he completed his PhD in communication and information systems under Guangcan Guo, the “Quantum Control” 973 chief scientist. His research interests include quantum communication devices and networking.