苹果的人工智能专家畅谈强化学习

本周二，苹果（Apple）人工智能研究部门的主任鲁斯兰·萨拉赫丁诺夫探讨了这项技术的部分局限性。不过，在这次《麻省理工科技评论》（MIT Technology Review）的会议上，对于他所在的神秘公司会如何将人工智能应用于Siri等产品，他回避了讨论。

去年10月加入苹果的萨拉赫丁诺夫表示，他对于人工智能中强化学习这个领域尤其有兴趣。利用这种方法，研究人员可以教电脑反复采取各种举动，找出最优的解决方案。例如，谷歌（Google）就用强化学习帮助其数据中心的计算机找到了最优的冷却和运转配置，从而提高了能源的利用效率。

萨拉赫丁诺夫还是卡内基梅隆大学（Carnegie Mellon）的助理教授。他说，这所大学最近在利用强化学习，训练计算机玩一款20世纪90年代的电子游戏《毁灭战士》（Doom）。计算机很快就学会了准确射击外星人，还发现回避动作可以躲开敌军的火力。然而，这些专业的《毁灭战士》计算机系统不太善于记忆迷宫场景等，这导致它们无法规划和设计策略。

萨拉赫丁诺夫的研究目标之一，就是开发能记住《毁灭战士》虚拟迷宫及几个特殊参照点，从而定位特定巨塔的人工智能软件。这款软件在游戏中会先查看火炬是红色还是绿色，因为火炬颜色的不同，意味着需要定位的巨塔颜色不同。

最后，软件学会了通过迷宫，抵达正确的巨塔。如果它走错了，也会原路返回迷宫找到正确的那个。萨拉赫丁诺夫表示，尤其值得注意的是，软件每次看到巨塔，都能回忆起开始看到的火炬颜色。

然而，他也表示，这种人工智能需要“长期的训练时间”，还要求强大的运算能力，因此很难大批量生产。他说：“目前来看，它还太脆弱了。”

萨拉赫丁诺夫另一个想要探索的领域，就是让人工智能软件更快地通过“少数案例和经历”进行学习。尽管他没有明说，不过他的想法应该能帮助苹果在“用更少时间做出更好产品”的竞争中取得优势。

一些人工智能的专家和分析师认为，苹果的人工智能技术比起谷歌或微软（Microsoft）等对手要逊色一筹。因为公司有着更严格的用户隐私条款，限制了能够用于训练计算机的数据量。如果苹果在训练计算机上使用了更少的数据，公司或许会在满足隐私要求的情况下，用媲美竞争对手的速度改进软件。（财富中文网）

作者：Jonathan Vanian

译者：严匡正

On Tuesday, Apple’s director of AI research, Ruslan Salakhutdinov, discussed some of those limitations. However, he steered clear during his talk at an MIT Technology Review conference of how his secretive company incorporates AI into its products like Siri.

Salakhutdinov, who joined Apple in October, said he is particularly interested in a type of AI known as reinforcement learning, which researchers use to teach computers to repeatedly take different actions to figure out the best possible result. Google (goog, +0.17%), for example, used reinforcement learning to help its computers find the best possible cooling and operating configurations in its data centers, thus making then more energy efficient.

Researchers at Carnegie Mellon, where Salakhutdinov is also an associate professor, recently used reinforcement learning to train computers to play the 1990's era video game Doom, Salakhutdinov explained. Computers learned to quickly and accurately shoot aliens while also discovering that ducking helps with avoiding enemy fire. However, these expert Doom computer systems are not very good at remembering things like the maze's layouts, which keeps them from planning and building strategies, he said.

Part of Salakhutdinov’s research involves creating AI-powered software that memorizes the layouts of virtual mazes in Doom and points of references in order to locate specific towers. During the game, the software first spots what's either a red or green torch, with the color of the torch corresponding to the color of the tower it needs to locate.

Eventually, the software learned to navigate the maze to reach the correct tower. When it discovered the wrong tower, the software backtracked through the maze to find the right one. What was especially noteworthy was that the software was able to recall the color of the torch each time it spotted a tower, he explained.

However, Salakhutdinov said this type of AI software takes “a long time to train” and that it requires enormous amounts of computing power, which makes it difficult to build at large scale. “Right now it’s very brittle,” Salakhutdinov said.

Another area Salakhutdinov wants to explore is teaching AI software to learn more quickly from “few examples and few experiences.” Although he did not mention it, his idea would benefit Apple in its race to create better products in less time.

Some AI experts and analysts believe Apple's AI technologies are inferior to competitors like Google or Microsoft because o听f the company's stricter user privacy rules, which limits the amount of data it can use to train its computers. If Apple used less data for computer training, it could perhaps satisfy its privacy requirements while still improving its software as quickly as rivals.