【转】A Search Engine for Your Memories


From The Atlantic: A girl explores a huge model of the brain at the Shanghai Science and Technology Museum, in 2003

An inventor at IBM has patented technology for a cognitive assistant that could learn all about you, then remind you of a name you can’t remember the moment you need to say it.

People are always forgetting names. That’s because, at least in part, names are arbitrary. A name, in and of itself, doesn’t offer much context. And contextual associations are a big part of how humans form and access memories—the way the smell of pine trees might stir a memory of summer camp, or how hearing the hook of “Sweet Caroline” might transport you momentarily to Fenway Park.

A truly personalized search engine could be so useful this way. Imagine being able to google the ordinary things that slip your mind: “Where did I leave my sunglasses?” or “Wait, what were we just talking about?”

People have, at times, lamented the way the Internet can be a ruiner of ordinary mysteries. Instead of idly wondering about something you know is there in your brain somewhere—Quick! Hum the theme song from Alf!—and letting it inexplicably pop back into your consciousness in time (or not), the answers are always at our fingertips. (Yup, that’s the one.)

If it were possible to dip into the human mind the way people seek and retrieve information online, a sense of curiosity would, I am sure, remain intact—it’d just be redirected. In some cases finding an answer means moving on to the next thing; in others, it means an opportunity to go even deeper.

“Human memory is not the same as computer memory,” said James Kozloski, an inventor at IBM who focuses on computational and applied neuroscience. “We don’t have pointers. We don’t have addresses where we can just look up the data we need.”

Kozloski wants to change that. He recently filed a patent for technology that, in the simplest terms, will help finish your sentences for you. Like autocomplete for your voice, the system is a model of human memory that could be embedded in a device and offer prompts when necessary. It would use a combination of surveillance, machine learning, and Bayesian inference—a kind of predictive modeling—to recognize when a person has forgotten something, then provide the missing information.

“The idea is quite simple,” Kozloski told me. “You monitor an individual’s context, whether it’s what they’re saying or what they’re doing … and you predict what comes next.”

The monitoring could be done in many of the ways that people are already using sensors today. It might involve Fitbit-like wearables; movement trackers like the ones smart thermostats use to determine when a person walks from one room to the next; and WiFi-connected microphones like the ones the newfangled Barbie dolls have so they can listen and reply to children. Which is to say, if you’re unsettled by the notion of a future in which omnipresent computers are watching you and listening to you: That future is already here.

A cognitive assistant would probably have to draw on short-range wireless technologies that could pair with sensors to figure out exactly what a person’s doing: distinguishing, for example, between the arm movement you make when you brush your teeth versus how your arm moves when you’re dicing garlic.

If you can get past the creepiness factor, a cognitive interface like the one Kozloski envisions could theoretically be useful for anyone, but he sees specific applications for people as they get older—and especially for those who suffer from diseases like Alzheimer’s. “The loss of ability to access memory in the moment is the beginning of the breakdown of normal cognitive function: the ability of individuals to interact with others, take care of themselves, clothe themselves, cook meals,” he said.

Such a system could help caregivers track how people are doing over time—are they forgetting important tasks more frequently?—and “perhaps prevent side effects of what are otherwise sort of innocuous episodes of forgetting,” Kozloski said. “Like getting confused, getting agitated, then putting myself at a greater risk.”

Imagine for example, if your cognitive assistant knew that when you dial a certain person’s phone number—your niece, let’s say—it should also remind you of the name of her husband. The system might also know that, because of the time of day when you’re calling, the husband is more likely to pick up the phone. Or that, by checking a calendar, it happens to be his birthday.

“All of that context becomes the basis for inference as to what name should be spoken when they pick up the phone,” Kozloski said.

That information—Paige’s husband’s name is Teddy. And it’s his birthday!— could come through an earpiece or mobile device. Or it could be part of a larger—and, frankly, more Hal 9000-esque—home integration. “It could be just the speakers in your house that are primed and ready to advise, literally waiting in the wings to assist,” Kozloski said.

Waiting, but not always jumping in. That’s key.

“It would be very annoying if it were continually interrupting you,” Kozloski said. So the system would use basic pause-detection, reinforced by machine learning as a way to understand the cadence and rhythm of a person’s speech and behavioral patterns. Once a cognitive assistant knows a person’s typical morning routine, for example, it can figure out whether something is amiss. Then it can decide whether and how to help.

“An individual’s behaviors are mapped to a unique activity such as ‘putting socks on,’ or ‘making the bed,’ or ‘brushing teeth,’ or ‘chopping onions,’” Kozloski said. “All of those events can form a sort of Markov network of probabilities where we can predict with some degree of certainty, and measure the likelihood [that what’s happening] isn’t normal for that person.”

The technology would also need human feedback, either from the person using it or a caregiver, to confirm and improve the accuracy of what the machine learns over time. “The opportunity to personalize this to individual cadence or idiosyncrasies,” Kozloski said “to allow a caregiver to augment that learning such that you can disambiguate whether an individual is confused or just sleepy. All of that is in the realm of the patent.”

Which means eventually, people may not even find themselves wishing they could Google their own memories. They won’t have to.

This entry was posted in 新闻动态 and tagged . Bookmark the permalink.

One Response to 【转】A Search Engine for Your Memories

  1. 王海鹏 says:

    1. 如果将记忆视为”大脑”最主要的功能体现之一,我想Memory这个函数的返回值应该是一个布尔值:记得or忘记,当然这也是大脑的神经网络底层活动过程的具体表现。 我认为在上文中“记忆”这个词应该指“记得”, 在人生中我们不断的获得输入(五感获得的信息),并基于与生俱来的神经元网络不断构建(通过思考和实践)新的神经元,这让我们“记得”一些内容。因此,我们“记得”是依靠“相互关联的神经元们”,这启发了我们:相互关联的神经元让我们“记得”一个相互关联的世界。
    2. 这似乎也是上文的思路(第二段:”And contextual associations are a big part of how humans form and access memories”)即我们基于“输入”和“上下文的关联”启动我们的记忆函数。我非常认同这种观点,同时我认为这种记忆过程证明了:通过事物之间的关联来认知世界正是人类与生俱来的认知世界的主要方式之一。
    3. 基于普通人只拥有有限额神经元集合(10^11个),遗忘这个过程也显得理所当然。正因为任何时刻的信息输入,遗忘和记忆在人的一生中不断重复。因此通过事物的关联来建立自己的认知系统是伴随我们一生的。
    4. 在互联网时代,各种新奇的科技发展迅速,我们获取信息的代价越来越小,效率越来越高,而探索事物的关联性无疑需要大量信息数据来模拟现实。因此,我们有理由相信:在这个时代我们有资本和条件以关联的角度看待世界,而第一步就是上文所说(“we can just look up the data we need”)
    5. 但是,我对上文中的部分举例有疑问。我认为多数案例所做的工作是:“让机器通过人的行为提供的输入,关联(猜)到更多的信息”,这样在一些特殊情况下(如:阿尔茨海默症患者意识不到自己需要的东西)可以帮助人“记起”一些事。这似乎是“帮助人记忆”的过程,但又像是“警告人应该做什么”,比如今天机器人提醒阿尔茨海默患者中午11点需要服药,这能帮助患者在第二天同时记起需要吃药这件事么?
    6. 总之,我非常赞同文章关于“关联性认知”的思路,并认为互联网时代是前所未有的尝试和实验“关联认知”的时代。但我认为“记忆”始终时人类最重要的活动之一,我们最需要的是帮助(比如启发)人们记忆,而不是代替人们记忆。