评论
打印
The central processing unit, or CPU, that's the key to making your home computer work is often likened to a brain,
中央处理器，或者说CPU，是让你的家用电脑工作的关键，通常将其比作大脑，
but the truth is it's nothing like the brains found in nature or in our skulls.
但事实上，它与自然界其他动物或我们人类头骨中的大脑完全不同。
CPUs are great at performing precise calculations with huge numbers,
CPU擅长用大量的数字进行精确计算，
but when it comes to learning and abstraction, the thinky meat between our ears has the CPU licked.
但是当涉及到学习和抽象时，我们耳朵之间会思考的那块肉可以打败CPU。
An emerging field of artificial intelligence called neuromorphic computing is attempting to mimic how the neurons in our own brains work,
一个叫“神经形态计算”的新兴人工智能领域正试图模拟我们大脑中神经元的工作方式，
and researchers from Intel and IBM are making true silicon brains a reality.
来自英特尔和IBM的研究人员正在使真正的硅脑成为现实。
Now, it's easy to get a little lost in the terminology here because another technology on the forefront of AI is called deep learning,
现在，我们很容易对这些术语摸不着头脑，因为人工智能前沿的另一项技术叫做深度学习，
and one of the most advanced approaches relies on something called a neural network.
最先进的方法之一是依靠一种叫做神经网络的东西。
Neural networks are a software approach that mimic how brains work.
神经网络是一种模拟大脑工作方式的软件。
A neural network changes when it's shown lots and lots of examples of what it's supposed to learn,
当一个神经网络显示出很多它应该学习的东西时，它就会发生变化，
but it may need to see thousands to millions of examples to achieve the desired results, like how to tell the difference between a chihuahua and a blueberry muffin.
但它可能需要看到数千到数百万个例子才能达到预期的效果，比如如何区分吉娃娃和蓝莓松饼。
Clearly that's not how we learn.
显然我们不是这样学习的。
I don't need to see millions of pictures of a dog before I know what a dog is.
在我知道什么是狗之前，我不需要看上百万张狗的照片。
But if you want to send me pictures of your dog I am on Twitter.
但如果你想把你的狗的照片发给我，这是我的推特账号。
So, to solve this, researchers from IBM and Intel are trying to mimic brains at a hardware level too.
为了解决这个问题，IBM和英特尔的研究人员正尝试在硬件层面上模拟大脑。
IBM revealed their brain inspired chip called TrueNorth in 2014, while Intel's chip called Loihi was introduced in 2017.
2014年，IBM公布了他们的大脑启发芯片TrueNorth，而英特尔的芯片Loihi则在2017年推出。
The two neuromorphic chips use the same silicon transistors commonly found in conventional chips,
这两个神经形态芯片使用的是传统芯片中常见的相同硅晶体管，
but they're arranged to interconnect more like neurons.
但它们的排列更像神经元的相互连接。
TrueNorth's one million neurons are connected by 256 million synapses,
TrueNorth的一百万个神经元由2.56亿个突触连接，
while Loihi's 130,000 neurons are each capable of communicating with thousands of others for a total of over 130 million synapses.
而Loihi的13万个神经元中的每一个都能与成千上万的其他神经元进行交流，总共超过1.3亿个突触。
TrueNorth and Loihi also combined into one chip two aspects of computers that are normally separate: memory and computation.
TrueNorth和Loihi还将计算机的两个方面（通常是分开的）组合成一个芯片：内存和计算。
In a typical computer like you have at home, the CPU handles computation and shuffles data back and forth from the Random Access Memory, or RAM.
在你们家里的普通计算机中，CPU处理计算并从随机存取存储器（RAM）来回洗牌数据。
But this separation slows things down and draws more power, and it's not how things work in our own brains.
但这种分离会减慢速度并消耗更多的电量，这不是我们大脑的运作模式。
In another drastic departure from standard chips, TrueNorth and Loihi do not use a clock to update information across the system in a synchronized manner.
另一个与标准芯片大相径庭的是，TrueNorth和Loihi不使用计时器来同步更新整个系统的信息。
Instead, the neurons in the chip fire independently, and the timing of these spikes of activity can be used as another way to encode information.
芯片中的神经元是独立工作的，这些活动峰值的时间可以作为另一种编码信息的方式。
All of these tweaks to how information is moved around means neuromorphic chips can learn quickly and use far less energy than a conventional CPU.
所有这些关于信息传递方式的调整意味着神经形态芯片可以比传统的CPU更快地学习和消耗更少的能量。
Best of all though, is the problems they can solve as a result of their novel design.
最棒的地方在于它们依靠新颖设计而可以解决的问题。
Problems like constraint satisfaction, where several solutions could exist but only one of them fits the constraints.
包括约束满足等可能存在多个解决方案的问题，但其中只有一个满足约束条件。
Think Sudoku puzzles.
想想数独游戏吧。
Neuromorphic computers can also be used for optimization tasks,
神经形态计算机也可以用于优化任务，
like the famous traveling salesman problem where finding the best route to take from millions of options can be very challenging, even for a supercomputer.
比如著名的旅行推销员问题，我们很难从数百万种选择中找到最佳方案，即便是用超级计算机来做也是如此。
Since Loihi is a research chip that was never intended for mass production, there aren't many of them for researchers to work with.
Loihi是一款从未打算大规模生产的研究型芯片，因此没有多少可供研究人员使用。
Still, Intel wired together 768 of them to create Pohoiki Springs, a computer that's the size of 5 servers and boasts 100 million neurons.
尽管如此，英特尔还是将其中的768个连接在一起，创建了Pohoiki Springs：一个有5台服务器大小、拥有1亿个神经元的计算机。
That's in league with the brain size of a small mammal.
这与小型哺乳动物的大脑大小相当。
And yet, despite its size and complexity, it needed under 500 watts of power to operate.
但尽管它有如此的规模和复杂性，其运行所需的电力只需要不到500瓦。
By contrast, the overkill gaming PC sitting next to me can use up to twice that much power, and it still isn't as "smart" as a squirrel.
相比之下，我旁边的这台超级强大的游戏电脑要消耗2倍的电力，而且它还没有松鼠那么“聪明”。
Neuromorphic computers are not poised to completely replace conventional ones any time soon.
神经形态计算机并不会在短期内完全取代传统计算机。
Remember that because this kind of hardware is just emerging, software that can make the best use of it needs time to develop.
请记住，这种硬件才刚刚出现，因此能够充分利用其效能的软件需要一段时间来开发。
Still, it's something to look forward to.
不过这还是值得期待的。
As the technology matures we'll be able to crack bigger and tougher problems that were previously beyond our grasp with our current CPU "brains."
随着技术的成熟，我们将能够破解更大、更棘手的问题，而这些问题在我们目前的CPU“大脑”中是无法解决的。
While our brains are more adaptable than a conventional CPU, our data processing speed is estimated to be a paltry 120 bits per second.
虽然我们的大脑比传统的CPU更具适应性，但我们的数据处理速度估计只有微不足道的每秒120位。
If you want to know more about neural networks, check out Maren's video on how robots teach themselves here!
如果你们想了解更多关于神经网络的知识，可以看看马伦的视频，它介绍了机器人是如何自学的！
If you like this video be sure to let us know in the comments, or subscribe! Then we know you really like us.
如果你们喜欢这个视频，请一定要在评论里留言或订阅！这样我们就知道你们是真的喜欢这些视频。
Thanks for watching and I'll see you next time on Seeker!
谢谢收看，下次《科学探秘之旅》再见！
