00:12

I study ants in the desert, in the tropical forest and in my kitchen, and in the hills around Silicon Valley where I live. I've recently realized that ants are using interactions differently in different environments, and that got me thinking that we could learn from this about other systems, like brains and data networks that we engineer,and even cancer.

00:41

So what all these systems have in common is that there's no central control. An ant colony consists of sterile female workers -- those are the ants you see walking around — and then one or more reproductive femaleswho just lay the eggs. They don't give any instructions. Even though they're called queens, they don't tell anybody what to do. So in an ant colony, there's no one in charge, and all systems like this without central control are regulated using very simple interactions. Ants interact using smell. They smell with their antennae,and they interact with their antennae, so when one ant touches another with its antennae, it can tell, for example, if the other ant is a nestmate and what task that other ant has been doing. So here you see a lot of ants moving around and interacting in a lab arena that's connected by tubes to two other arenas. So when one ant meets another, it doesn't matter which ant it meets, and they're actually not transmitting any kind of complicated signal or message. All that matters to the ant is the rate at which it meets other ants. And all of these interactions, taken together, produce a network. So this is the network of the ants that you just saw moving around in the arena, and it's this constantly shifting network that produces the behavior of the colony,like whether all the ants are hiding inside the nest, or how many are going out to forage. A brain actually works in the same way, but what's great about ants is that you can see the whole network as it happens.

02:23

There are more than 12,000 species of ants, in every conceivable environment, and they're using interactions differently to meet different environmental challenges. So one important environmental challenge that every system has to deal with is operating costs, just what it takes to run the system. And another environmental challenge is resources, finding them and collecting them. In the desert, operating costs are high because water is scarce, and the seed-eating ants that I study in the desert have to spend water to get water. So an ant outside foraging, searching for seeds in the hot sun, just loses water into the air. But the colony gets its water by metabolizing the fats out of the seeds that they eat. So in this environment, interactions are used to activate foraging. An outgoing forager doesn't go out unless it gets enough interactions with returning foragers, and what you see are the returning foragers going into the tunnel, into the nest, and meeting outgoing foragers on their way out. This makes sense for the ant colony, because the more food there is out there, the more quickly the foragers find it, the faster they come back, and the more foragers they send out.The system works to stay stopped, unless something positive happens.

03:39

So interactions function to activate foragers. And we've been studying the evolution of this system. First of all, there's variation. It turns out that colonies are different. On dry days, some colonies forage less, so colonies are different in how they manage this trade-off between spending water to search for seeds and getting water back in the form of seeds. And we're trying to understand why some colonies forage less than others by thinking about ants as neurons, using models from neuroscience. So just as a neuron adds up its stimulationfrom other neurons to decide whether to fire, an ant adds up its stimulation from other ants to decide whether to forage. And what we're looking for is whether there might be small differences among colonies in how many interactions each ant needs before it's willing to go out and forage, because a colony like that would forage less.

04:32

And this raises an analogous question about brains. We talk about the brain, but of course every brain is slightly different, and maybe there are some individuals or some conditions in which the electrical properties of neurons are such that they require more stimulus to fire, and that would lead to differences in brain function.

04:53

So in order to ask evolutionary questions, we need to know about reproductive success. This is a map of the study site where I have been tracking this population of harvester ant colonies for 28 years, which is about as long as a colony lives. Each symbol is a colony, and the size of the symbol is how many offspring it had,because we were able to use genetic variation to match up parent and offspring colonies, that is, to figure out which colonies were founded by a daughter queen produced by which parent colony. And this was amazing for me, after all these years, to find out, for example, that colony 154, whom I've known well for many years, is a great-grandmother. Here's her daughter colony, here's her granddaughter colony, and these are her great-granddaughter colonies. And by doing this, I was able to learn that offspring colonies resemble parent colonies in their decisions about which days are so hot that they don't forage, and the offspring of parent colonies live so far from each other that the ants never meet, so the ants of the offspring colony can't be learning this from the parent colony. And so our next step is to look for the genetic variation underlying this resemblance.

06:07

So then I was able to ask, okay, who's doing better? Over the time of the study, and especially in the past 10 years, there's been a very severe and deepening drought in the Southwestern U.S., and it turns out that the colonies that conserve water, that stay in when it's really hot outside, and thus sacrifice getting as much food as possible, are the ones more likely to have offspring colonies. So all this time, I thought that colony 154 was a loser, because on really dry days, there'd be just this trickle of foraging, while the other colonies were outforaging, getting lots of food, but in fact, colony 154 is a huge success. She's a matriarch. She's one of the rare great-grandmothers on the site. To my knowledge, this is the first time that we've been able to track the ongoing evolution of collective behavior in a natural population of animals and find out what's actually working best.

07:04

Now, the Internet uses an algorithm to regulate the flow of data that's very similar to the one that the harvester ants are using to regulate the flow of foragers. And guess what we call this analogy? The anternet is coming.(Applause) So data doesn't leave the source computer unless it gets a signal that there's enough bandwidthfor it to travel on. In the early days of the Internet, when operating costs were really high and it was really important not to lose any data, then the system was set up for interactions to activate the flow of data. It's interesting that the ants are using an algorithm that's so similar to the one that we recently invented, but this is only one of a handful of ant algorithms that we know about, and ants have had 130 million years to evolve a lot of good ones, and I think it's very likely that some of the other 12,000 species are going to have interesting algorithms for data networks that we haven't even thought of yet.

08:10

So what happens when operating costs are low? Operating costs are low in the tropics, because it's very humid, and it's easy for the ants to be outside walking around. But the ants are so abundant and diverse in the tropics that there's a lot of competition. Whatever resource one species is using, another species is likely to be using that at the same time. So in this environment, interactions are used in the opposite way. The system keeps going unless something negative happens, and one species that I study makes circuits in the trees of foraging ants going from the nest to a food source and back, just round and round, unless something negative happens, like an interaction with ants of another species. So here's an example of ant security. In the middle, there's an ant plugging the nest entrance with its head in response to interactions with another species. Those are the little ones running around with their abdomens up in the air. But as soon as the threat is passed, the entrance is open again, and maybe there are situations in computer security where operating costs are low enough that we could just block access temporarily in response to an immediate threat, and then open it again, instead of trying to build a permanent firewall or fortress.

09:33

So another environmental challenge that all systems have to deal with is resources, finding and collecting them. And to do this, ants solve the problem of collective search, and this is a problem that's of great interestright now in robotics, because we've understood that, rather than sending a single, sophisticated, expensive robot out to explore another planet or to search a burning building, that instead, it may be more effective to get a group of cheaper robots exchanging only minimal information, and that's the way that ants do it. So the invasive Argentine ant makes expandable search networks. They're good at dealing with the main problem of collective search, which is the trade-off between searching very thoroughly and covering a lot of ground. And what they do is, when there are many ants in a small space, then each one can search very thoroughlybecause there will be another ant nearby searching over there, but when there are a few ants in a large space,then they need to stretch out their paths to cover more ground. I think they use interactions to assess density,so when they're really crowded, they meet more often, and they search more thoroughly. Different ant species must use different algorithms, because they've evolved to deal with different resources, and it could be really useful to know about this, and so we recently asked ants to solve the collective search problem in the extreme environment of microgravity in the International Space Station. When I first saw this picture, I thought, Oh no, they've mounted the habitat vertically, but then I realized that, of course, it doesn't matter. So the idea here is that the ants are working so hard to hang on to the wall or the floor or whatever you call it that they're less likely to interact, and so the relationship between how crowded they are and how often they meet would be messed up. We're still analyzing the data. I don't have the results yet. But it would be interesting to know how other species solve this problem in different environments on Earth, and so we're setting up a program to encourage kids around the world to try this experiment with different species. It's very simple. It can be done with cheap materials. And that way, we could make a global map of ant collective search algorithms. And I think it's pretty likely that the invasive species, the ones that come into our buildings, are going to be really good at this, because they're in your kitchen because they're really good at finding food and water.

12:09

So the most familiar resource for ants is a picnic, and this is a clustered resource. When there's one piece of fruit, there's likely to be another piece of fruit nearby, and the ants that specialize on clustered resources use interactions for recruitment. So when one ant meets another, or when it meets a chemical deposited on the ground by another, then it changes direction to follow in the direction of the interaction, and that's how you get the trail of ants sharing your picnic.

12:36

Now this is a place where I think we might be able to learn something from ants about cancer. I mean, first, it's obvious that we could do a lot to prevent cancer by not allowing people to spread around or sell the toxins that promote the evolution of cancer in our bodies, but I don't think the ants can help us much with thisbecause ants never poison their own colonies. But we might be able to learn something from ants about treating cancer. There are many different kinds of cancer. Each one originates in a particular part of the body,and then some kinds of cancer will spread or metastasize to particular other tissues where they must be getting resources that they need. So if you think from the perspective of early metastatic cancer cells as they're out searching around for the resources that they need, if those resources are clustered, they're likely to use interactions for recruitment, and if we can figure out how cancer cells are recruiting, then maybe we could set traps to catch them before they become established.

13:37

So ants are using interactions in different ways in a huge variety of environments, and we could learn from thisabout other systems that operate without central control. Using only simple interactions, ant colonies have been performing amazing feats for more than 130 million years. We have a lot to learn from them.

13:58

Thank you.

14:01

(Applause)

00:12

我研究各种蚂蚁 沙漠中的、热带雨林里的、 我厨房里的蚂蚁， 以及我硅谷的家周边山上的蚂蚁。 最近我注意到蚂蚁在不同的环境下 交互方式也是不同的， 这让我想到或许我们能从中学到些什么 用到其它系统上。 例如大脑结构或者我们的数据网络 甚至是癌症。 这些系统的共同点在于 没有一个中央控制结构。 蚁群的工蚁由不育的雌性构成— 工蚁就是你能看到的蚂蚁— 而能够生育的雌性蚂蚁（蚁后） 只负责产卵。 蚁后不会发号指令。 虽然它们叫做蚁后， 但是它们不会指挥其它工蚁。 所以任何蚁群都没有一个最高负责人， 所有这些系统都是没有中央控制的， 仅仅通过简单的交互方式进行运作. 蚂蚁的交互是通过嗅觉进行的. 它们用触角（antennae）去嗅. 用触角来交流。 所以当一只蚂蚁的触角碰到另一只蚂蚁的触角 它就知道另一个蚂蚁 是不是同一个巢穴的 以及这只蚂蚁正要做什么事情. 现在你看到的这个蚂蚁的活动场所 通过玻璃管子跟另外两个场所连接着 蚂蚁在这些活动场所里走来走去. 当一只蚂蚁遇到了另外一支蚂蚁, 遇见的是哪只蚂蚁并不重要, 它们也没通过触角传递 任何复杂的信号或消息. 唯一传递的是两只蚂蚁 相互遇见的频率. 这些交互信息汇总起来后, 我们就得到了一个网络. 这就是刚才你看到的蚂蚁 四处移动之后生成的网络图, 正是这张不断变化中的网络, 塑造了这个蚁群的行为, 像是有多少蚂蚁躲在巢穴里, 多少蚂蚁出去寻食之类的信息. 大脑差不多也是这么工作的, 相比起来观察蚂蚁吸引人的地方之一, 是你可以看到整个网路是如何运作的. 蚂蚁的种类超过一万两千种, 你能想象到的环境里都有蚂蚁存在, 而且不同环境下的蚁群会使用 不同的交流方式以适应环境特点. 例如不同环境下不同蚁群 普遍面临的问题之一 是如何控制"运营开支", 即需要花多大成本 才能生存下来. 另一个环境带来的挑战是, 如何去搜寻和收集资源. 在沙漠中, 水非常的稀少, 所以运营开支很大, 我研究的一种生活在沙漠中以植物种子为食的蚂蚁 寻找水源的同时需要消耗水. 所以当一只蚂蚁外出觅食的时候, 在火辣辣的太阳底下找种子的时候, 它体内的水分会被蒸发. 而蚁群 可以通过消化种子富含的脂肪 产生需要的水. 所以在这种环境下, 蚂蚁之间的交互 主要用来决定是否外出觅食. 一个准备外出的觅食者不会轻易外出, 除非得到了足够的归巢的觅食者的反馈, 你现在看到的是回来的觅食者, 在通过蚁巢的管道进入蚁穴时, 跟沿路准备外外出的蚂蚁进行交流. 这对蚁群来说很重要, 因为外面的食物越多, 觅食的蚂蚁找到食物的速度越快, 它们回来的就更快, 那么就会有更多的蚂蚁出去觅食. 这个系统默认的行为是按兵不动, 除非看到了足够的好处. 所以在这里交互是为了决定是否出去觅食. 我们已经研究这种系统演化有一段时间了. 首先, 这种演化各不相同. 不同的蚁群的行为是不一样的. 在旱季, 有些蚁群觅食的少, 不同蚁群之间的差异 就体现在它们如何做权衡 如何在消耗更多水分去寻找食物 以及获得更多食物和水之间权衡 我们尝试将蚁群 类比成神经细胞组织基于脑神经科学的相关理论 来理解蚁群觅食行为的差异。 所以就像是一个神经元是否触发， 取决于相连的神经元触发强度之和， 蚂蚁的行为也由其它蚂蚁决定， 是否要出去觅食。 于是我们就希望能够找到 觅食行为存在差异的蚁群之间 是否蚂蚁在觅食前交互的其它蚂蚁数量 也是存在对应差异的。 因为像那样的蚁群会更少外出觅食。 这个问题也可以用大脑来进行类比。 我们提到的大脑 当然也是每个大脑都有些许不同的 肯定有一些个体在某些环境下 他们的神经元的电特性决定了 需要接受更多的刺激才会激发。 而这会导致脑的功能差异。 而为了解答之前系统演化的问题， 我们首先需要研究下后代繁殖率。 这张图显示的是我的研究站附近的蚁群图 我在这个地方研究收获蚂蚁（一种西方蚁） 种群演化已经超过28年了。 这大概也是一个种群能够延续的时间。 每一个圆圈都表示一个种群， 圆圈的大小表示后代的规模， 我们可以通过基因变化分析(genetic variation) 来确认种群之间的父子关系， 也就是能够确认每个蚁群 里面的蚁后来自于 哪个父代蚁群。 研究这么多年之后我有了一些迷人的发现，例如，154号种群， 我研究很多年的这个， 算是祖母级别的。 这是她的女儿种群， 这是她的孙女种群， 这是重孙女种群。 分析这些种群使我能够 发现后代种群（的多少）体现了 父代种群在炎热天气下 觅食的策略差异， 而且考虑到父代种群 与后代种群之间距离很远，不可能遇见， 所以后代种群中的蚂蚁 不会从父代种群那里学习到什么。 于是第二步就是看看 这种相似性的基因学变异根源。 然后我就可以提出这个问题：哪群蚂蚁的策略更好？ 在研究进行中的那些年里， 尤其是最近的十年， 实验所在的美国西南部 经历了非常严重和持久的干旱， 结果是那些更注重保持水分的蚁群， 那些大热天不出门的蚁群， 也就是那些失去了更多觅食机会的蚁群， 反而是更有可能有后代蚁群的。 我曾经一度认为154号种群 是进化的失败者， 因为在旱季， 它们很少出去觅食， 反之其它的种群 会出去寻找更多的食物， 但是结果是，154号种群非常的成功。 她是事实上的统领。她是这个研究点非常少见的有重孙后代的蚁群 就我所知，这还是第一次 我们人类能够追踪到 自然界中野生生物群体的 集体行为进化 以及找到最适合环境的生存方式. 现在, 互联网使用的算法 用来分配数据流动的算法 与这些蚂蚁使用的算法 即如何安排工蚁外出觅食的算法 非常相似. 你们猜我们如何称呼这种相似性? 蚁群互联网(Anternet)的到来. (掌声) 所以发送数据的电脑 在得到信号确认带宽足够之前 不会将数据发送出去. 在互联网的早期, 发送和接收数据的成本非常高, 所以任何形式的数据丢失都是不可以接受的, 所以网络系统被设计利用相互之间的交互 来决定何时发送数据. 发现蚂蚁跟我们人类最近才发明的算法 有这么大的相似性是很叫人惊喜的, 而且现在我们只发现了蚂蚁使用的算法中 一小部分的算法, 蚂蚁已经有了1.3亿年的历史 已经演化出很多好的算法, 因此我相信有可能 另外尚未研究的1.2万蚂蚁种类中 也有很多有意思的算法, 可以用于数据网络 这些算法甚至超过了我们的想象. 例如, 当运营成本很低的时候呢? 热带雨林里, 蚁群觅食的成本很低, 因为那里非常的湿润, 对于蚁群来说 外出觅食也非常容易. 但是蚂蚁的种类是如此的繁多 数量也非常庞大 因此蚂蚁之间的竞争非常激烈. 一个蚁群需要用到的任何资源 基本上都有竞争者 与之争夺. 所以在这样的环境下, 相互接触的用途 完全反了过来. 蚁群的系统不断的扩张, 直到一些不好的事情发生, 我研究的一种蚁群会在丛林里 构建自己的觅食网络, 在蚁穴和食物时间不断的来回, 一圈一圈的觅食, 直到一些不好的事情发生, 例如遇到了 别的种类的蚂蚁. 这是蚂蚁安防的一个例子. 中间的位置, 一只蚂蚁 在跟另外的种群的蚂蚁触碰了触角之后 将蚁穴的入口用自己的头挡住了. 这些小的、腹部朝上的蚂蚁 正在这周围走动. 但是一旦危险解除, 入口就会重新开启, 或许我们也可以联想到 在计算机安全领域 这个领域的运营成本也低到 我们可以临时的中断网络访问 以应对临时的威胁, 稍后继续开放, 而不是现在的做法 尝试构造一个永久的防火墙. 另一个环境带来的挑战 所有的蚁群系统都需要面对的 是如何寻找和搜集资源. 蚁群为了解决这个问题, 采用了 集体搜索(collective search)的方法, 而这个问题现在已经引起了 机器人研究人员的极大兴趣, 因为我们都知道, 与其用一个单一的 复杂且昂贵的机器人 去探索另外的星球 或去火场搜救, 或许有更好的方式 就是造一堆便宜的机器人 相互之间仅仅交换简单的信息, 就像是蚂蚁所做的那样. 这种外来的阿根廷蚂蚁 很擅长扩大自己的搜索网络. 它们非常善于解决集体搜索中的 主要问题, 即如何在两个不同的目标之间权衡 既要能够搜索的彻底 又要搜索的范围广. 它们是这么做的, 当搜索空间小而蚂蚁很多时, 它们会搜寻的非常彻底因为它们知道临近的区域 有别的蚂蚁在搜索, 但是当搜索面积很大 且蚂蚁很少时, 它们会扩张自己的搜索路径 去覆盖更大的面积. 我想它们之间的接触主要交换的是蚂蚁的密度信息, 当它们的密度很大时, 它们碰见的就越多, 搜寻的也就越仔细. 不同种类的蚂蚁使用的算法应该是不同的, 因为随着一代代的演化 它们需要的资源不同. 知道这些差异真的很有用. 所以最近我们把蚂蚁 放在微重力的极端环境中 希望能够帮助 国际空间站 解决集体搜索的难题. 当我第一次看到这张照片, 我想, 呀, 他们把蚁穴竖起来放着了, 但是马上意识到, 其实横竖都一样的. 这个实验的想法是 蚂蚁要花大力气把自己挂在墙上 或者也可以说是地板上, 你怎么看都行 这样它们就没有精力去交互了,所以关于蚂蚁密度的信息 以及它们相互遇见的频率 都会乱掉. 我们还在分析这些数据. 我还没有结论. 但如果我们能够知道地球上的 其它物种如何解决此类问题 这一定非常的有意思, 所以我们创建了一个活动 鼓励全世界的小朋友们 用不同的蚂蚁种类重复我们的实验. 非常简单. 做起来也不需要多少成本. 这样, 我们就能够绘制一张 蚂蚁集体搜索算法的"世界地图". 我想那些外来的蚂蚁种类, 那些混进我们大楼的蚂蚁, 对于集体搜索非常在行, 因为它们已经跑到你的厨房 非常地善于找到食物和水. 对于蚂蚁而言最为相似的资源 是野餐的地方, 是一个集中的资源.当一块水果掉在地上, 周围很可能还有更多的水果渣, 因此生活在集中资源多的地方的蚂蚁 通过相互接触来召集伙伴. 所以当一只蚂蚁遇见另一只蚂蚁, 或是另一只蚂蚁沿路留下的 化学气味, 然后它就会改变自己的方向 冲着接触方提供的方向去搜寻 这就是为什么能够有一只蚂蚁大军 与你分享野餐的原因. 现在, 我觉得我们或许可以 从蚂蚁身上获得治疗癌症的一些启发. 我是说, 首先, 我们可以做很多事情 来阻止癌症 例如禁止有人向其他人销售 可能增加我们身体患癌症风险的 有毒有害商品, 但是我不认为在这点上蚂蚁能够帮助我们什么, 因为它们从来不会毒害同类. 但是我们或许可以从蚂蚁那里学到一些方法 来治疗癌症. 癌症有很多不同的种类. 每一种癌症一开始都附着在身体的特定部位. 然后一些类型的癌症(癌细胞) 会扩散或传播到其它特定的组织结构中 它们需要在那里获得自己需要的资源. 现在如果你从这个角度 去看待早期癌细胞 它们也是在体内搜寻 寻找他们需要的资源, 如果这些资源是集中的, 那么它们很可能通过相互接触来召唤更多的癌细胞, 那么如果我们能够破解癌细胞相互召唤的机制 我们或许就能够设置陷阱 在癌细胞聚集之前捕获它们. 所以蚂蚁在不同的环境下 使用了完全不同的交互算法.我们能够从中学习 并将结果用于那些没有 中央控制的系统. 仅仅通过简单的接触, 蚂蚁已经创造了 长达1.3亿年的伟大历史. 我们还有很多需要向它们学习. 感谢大家. (掌声)