《财富》专访人工智能大牛杨立昆：“人工智能仍然缺乏常识！”

Facebook的首席人工智能科学家杨立昆（Yann LeCun）曾推动深度学习崛起，成为谷歌和亚马逊之类巨头纷纷应用的尖端人工智能技术，可迅速翻译并识别图片中的物品。

深度学习的核心是被称为神经网络的软件，可过滤海量数据，从而比人类更快地掌握数据中的模式。但该项技术需要巨大的计算能力，促使英特尔和硬件初创公司等半导体制造商努力设计全新的计算机芯片，降低能耗的同时提升一些人工智能计算任务效率。

星期一，杨立昆在旧金山举办的国际固态电路会议上发表新研究论文，概述他对未来人工智能的看法，着重关注芯片和硬件的发展前景。

以下是他谈到的几个要点：

1. 从翻译语言到管理内容

虽然Facebook、谷歌和微软等公司都打着降低能耗的口号研究专门的电脑芯片，不过杨立昆直接挑明了此类创新为何重要，因为新款芯片可以帮助各公司在自家数据中心应用更多神经网络。

因此，在线语音翻译等任务可能进一步完善，实现秒完成。与此同时，人工智能系统可以分析视频里的每一帧并识别人或物体，不仅仅是识别静态图像，由此大大提高准确性。

杨立昆还认为，使用性能更佳的计算机芯片可以提升内容管控效率，比如扫描文本中的攻击性语言或虚假新闻。对于努力删除平台上不良宣传或滥用行为的Facebook来说，相关技术进步越快越好。

2. 到处是“更智能”吸尘器和割草机的世界

杨立昆也在密切关注可供日常电器使用的电脑芯片，例如可以安装在吸尘器和割草机上的新产品。想象一下未来配置神经网络的割草机，可以轻松识别杂草和玫瑰，他解释说。

杨立昆还设想今后研发出性能更好的移动计算芯片，可以直接在移动设备上运行神经网络，不必再将信息传回数据中心计算。一些智能手机已经内置人工智能功能，比方说识别用户面部解锁设备，但如果想完成更复杂的任务，更先进的芯片必不可少。

他表示，人工智能的另一障碍是现在的电池续航能力。人工智能技术耗电量比较大，意味着一些较小的设备上使用人工智能会比较受限制。

3. 让电脑懂一些常识

尽管深度学习领域进步显著，但电脑仍然缺乏常识。现在的电脑要浏览数千张大象的照片才能在其他照片里独立识别出来。

相比之下，儿童只要对动物有了基本的理解，就可以迅速认识各种大象。即使增加认识的难度，孩子们还是能推断出大象只是一种体型巨大的动物。

杨立昆认为，人们终将开发出新型神经网络，可以通过筛选大量数据获得常识。过程类似于先传授基本技术，以后可用来参考，就像百科全书一样。然后，人工智能从业者可以进一步训练神经网络实现识别，并执行比现在更高级的任务。

但只有更先进的计算机芯片才可能实现，杨立昆希望尽快研发成功。（财富中文网）

译者：Pessy

审校：夏林

Facebook’s chief artificial intelligence scientist Yann LeCun helped spearhead the rise of deep learning, the cutting-edge AI technology used by companies like Google and Amazon to quickly translate languages and identify objects in photos.

At the core of deep learning is software called a neural network, which sifts through enormous amounts of data so that it can notice patterns more quickly than humans. But this technology requires tremendous computing power, prompting semiconductor makers like Intel and hardware startups to explore radical new computer chip designs for the job that gobble less energy and improve the efficiency of certain AI computational tasks.

LeCun present a new research paper on Monday at the International Solid State Circuits Conference in San Francisco that will outline his vision for AI’s future. In particular, he’ll focus on how the chips and hardware that makes it possible must evolve.

Here are a few highlights from his talk:

1. From translating languages to policing content

Although companies like Facebook, Google, and Microsoft are exploring specialized computer chips that reduce energy consumption, LeCun is blunt about why such innovation is important—new computer chips will allow companies to use even more neural nets inside their data centers than what’s possible today.

As a result, tasks like online speech translation could be supercharged so that they could be done in real time. Meanwhile, AI systems would be able to analyze every frame in a video in an effort to identify people or objects rather than just a few stills—thereby significantly boosting accuracy.

LeCun also believes that content moderation, like scanning text for offensive language or fake news, could be improved using better computer chips. For a company like Facebook that struggles with deleting propaganda and abusive behavior from its service, those advancements couldn’t come soon enough.

2. A world of “smarter” vacuum cleaners and lawnmowers

One trend LeCun is closely watching is computer chips that can fit in everyday devices like vacuum cleaners and lawnmowers. Imagine a futuristic lawnmower loaded with neural networks that could recognize the difference between weeds and garden roses, he explains.

LeCun also envisions even more sophisticated mobile computing chips that can run neural networks directly on the devices themselves rather than having to send information back to data centers for processing. Already, some smartphones are designed with AI built in that can recognize a user’s face to unlock the device, but improved computer chips will be necessary for more advanced tasks.

Another hurdle to AI are today’s batteries, he says. The technology eats a lot of energy, which means that using AI on some smaller devices is limited.

3. Giving computers some common sense

Despite advances in deep learning, computers still lack common sense. They would need to review thousands of images of an elephant to independently identify them in other photos.

In contrast, children quickly recognize elephants because they have a basic understanding about the animals. If challenged, they can extrapolate that an elephant is merely a different kind of animal—albeit a really big one.

LeCun believes that new kinds of neural networks will eventually be developed that gain common sense by sifting through a smorgasbord of data. It would be akin to teaching the technology basic facts that it can later reference, like an encyclopedia. AI practitioners could then refine these neural networks by further training them to recognize and carry out more advanced tasks than modern versions.

But it would only be possible using more powerful computer chips—ones that LeCun hopes are just around the corner.