ADM-201 dump PMP dumps pdf SSCP exam materials CBAP exam sample questions

确认AI程序的关键任务 – 译学馆
未登陆,请登陆后再发表信息
最新评论 (0)
播放视频

确认AI程序的关键任务

Verifying Mission-Critical AI Programs | Two Minute Papers #179

亲爱的学霸们
Dear Fellow Scholars,
这里是Károly Zsolnai Fehér带来的两分钟论文
this is Two Minute Papers with Károly Zsolnai-Fehér.
这篇论文不像之前的两分钟论文那样
This paper does not contain the usual fireworks
具有很大的视觉冲击力
that you’re used to in Two Minute Papers,
但我认为这是一个非常重要的 所有人都应该知道的事
but I feel that this is a very important story that needs to be told to everyone.
在计算机科学中 我们会遇到很多有趣的问题
In computer science, we encounter many interesting problems,
比如在一个城市给定的两条街道间找出最短路线
like finding the shortest path between two given streets in a city,
或者计算桥梁的稳固程度
or measuring the stability of a bridge.
直到几年前 这些问题几乎无一例外都依靠
Up until a few years ago, these were almost exclusively solved
传统的 手动的方法去解决
by traditional, handcrafted techniques.
意思是 科学家们手动设计出一套技术
This means a class of techniques that were designed by hand,
通常仅仅只是针对手头上现有的问题
by scientists and are often specific to the problem we have at hand.
不同的问题 不同的算法
Different problem, different algorithm.
然而 几年前
And, fast forward to a few years ago,
我们见证了神经网络和学习算法惊人的复兴
we witnessed an amazing resurgence of neural networks and learning algorithms.
很多以前认为不能解决的问题
Many problems that were previously thought to be unsolvable,
逐个被击破了
crumbled quickly one after another.
很显然 人工智能的时代已经来临了
Now it is clear that the age of AI is coming,
而且可以确定的是 对人工智能的一些可能的应用
and clearly, there are possible applications of it
我们需要多加小心
that we need to be very cautious with.
由于我们手动设计这些传统的技术
Since we design these traditional techniques by hand,
失败的案例通常都是已知的
the failure cases are often known
因为这些算法足够简单
because these algorithms are simple enough
我们可以查看内部然后做出合理的假设
that we can look under the hood and make reasonable assumptions.
但深度神经网络不是这样
This is not the case with deep neural networks.
我们知道在一些情况下 神经网络是不可靠的
We know that in some cases, neural networks are unreliable.
但是要辩识出这些失败的情况是非常困难的
But it is remarkably hard to identify these failure cases.
举个例子 之前我们讨论过名叫pix2pix的技术
For instance, earlier, we talked about this technique by the name pix2pix
我们可以画一幅猫的草图
where we could make a crude drawing of a cat
然后pix2pix就可以把它转换成真的照片
and it would translate it to a real image.
在很多情况下它都非常有效
It worked spectacularly in many cases,
但是推特上也充满了非常有趣的转换失败的例子
but twitter was also full of examples with really amusing failure cases.
除了不可靠性 我们还有一个更大的问题
Beyond the unreliability, we have a much bigger problem.
那就是 对抗样本
And that problem is adversarial examples.
在早期的视频里 我们讨论过一个对抗性算法
In an earlier episode, we discussed an adversarial algorithm,
在一个好玩的案例里
where in an amusing example,
他们给这幅图加上了一些几乎不可察觉的噪点
they added a tiny bit of barely perceptible noise to this image,
结果让深度神经网络错误地把公交车识别为了鸵鸟
to make the deep neural network misidentify a bus for an ostrich.
我们甚至可以训练一个新的神经网络
We can even train a new neural network
专门调整来突破现有神经网络的问题
that is specifically tailored to break the one we have,
加大它攻击目标的可能性
opening up the possibility of targeted attacks against it.
为了缓解这个问题
To alleviate this problem, it is always a good idea to make sure
确保这些神经网络也进行了对抗性输入的训练往往是个好办法
that these neural networks are also trained on adversarial inputs as well.
但是我们怎么才能知道到底还有多少其他对抗样本存在
But how do we know how many possible other adversarial examples exist
但是我们还没找到的呢?
that we haven’t found yet?
该论文讨论了一种检验神经网络重要性能的方式
The paper discusses a way of verifying important properties of neural networks.
比如 它可以测量这样一个神经网络的对抗鲁棒性
For instance, it can measure the adversarial robustness of such a network,
这超级有用
and this is super useful,
因为它可以让我们知道是否存在
because it gives us information whether there are possible forged inputs
可能破坏学习系统的伪造样本输入
that could break our learning systems.
该篇论文还包含一个很棒的小实验
The paper also contains a nice little experiment
是关于机载防撞系统的
with airborne collision avoidance systems.
它的目标是避免商用飞机在半空中相撞
The goal here is avoiding midair collisions
但是要把警报数量降到最小
between commercial aircrafts while minimizing the number of alerts.
作为一个小规模的思维实验
As a small-scale thought experiment,
我们可以训练一个神经网络来代替现有的系统
we can train a neural network to replace an existing system,
但是在这个案例里 这样一个神经网络必须是被检验过的
but in this case, such a neural network would have to be verified.
现在终于有可能性了
And it is now finally a possibility.
但是 不出错 不意味着
Now, make no mistake, this does not mean
在这个行业部署的各种飞行安全系统
that there are any sort of aircraft safety systems deployed in the industry
都依赖神经网络
that are relying on neural networks.
不不不 绝对不是
No no no, absolutely not.
这是一个小规模的“如果……”类型的实验
This is a small-scale “what if” kind of experiment
这可能是迈向真正激动人心的未来的第一步
that may prove to be a first step towards something really exciting.
这是一篇惊人的研究
This is one of those incredible papers that,
虽然可能没有以往的视觉冲击力
even without the usual visual fireworks,
但是让我感觉自己是未来的一部分
makes me feel that I am a part of the future.
这是迈向未来的一步
This is a step towards a future where we can prove
未来我们能够证明一个学习算法可以在任务的关键性系统中起作用
that a learning algorithm is guaranteed to work in mission critical systems.
我还要提醒一点
I would also like to note that
即使这一集不会在网络上大量传播
even if this episode is not meant to go viral on the internet,
这仍然是一个重要的新闻
it is still an important story to be told.
通常来说 制作这样一个视频是一种经济自杀的行为
Normally, creating videos like this would be a financial suicide,
但是我们完全不会因此受损
but we’re not hurt by this at all
因为我们在众筹网站上得到了你们的大力支持
because we get stable support from you on Patreon.
这就是它的全部意义
And that’s what it is all about –
——少担心点击量 多花点时间来探讨真正重要的问题
worrying less about views and spending more time talking about what’s really important.
这非常棒
Absolutely amazing.
感谢收看和你们的大力支持
Thanks for watching and for your generous support,
我们下期再见!
and I’ll see you next time!
[音乐播放中]
[Music Playing]

发表评论

译制信息
视频概述

深度神经网络成为近年来计算机发展的热点,但它仍然有很多难以解决的问题,如何确认一个AI学习程序的准确率高到可以实际应用?现在,这篇论文或许迈出了重要一步。

听录译者

收集自网络

翻译译者

吾家黄姑娘

审核员

审核团O

视频来源

https://www.youtube.com/watch?v=_DN2rzHkpZE

相关推荐