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

人工智能与神经任务规划 – 译学馆
未登录,请登录后再发表信息
最新评论 (0)
播放视频

人工智能与神经任务规划

Generalizing AI With Neural Task Programming | Two Minute Papers #206

亲爱的学术粉丝们 这里是
Dear Fellow Scholars, this is Two Minute Papers
Károly Zsolnai-Fehér的两分钟论文节目
with Károly Zsolnai-Fehér.
机器学习研究的”圣杯”问题之一
One of the holy grail problems of machine learning research
是实现人工智能 简而言之就是AGI
is to achieve artificial general intelligence, AGI in short.
深蓝已能打败国际象棋天才卡斯帕罗夫
Deep Blue was able to defeat the genius Kasparov in Chess,
但它无法告诉我们现在几点
but it was unable to tell us what the time was.
这种算法我们通常称为
Algorithms of this type we often refer to
弱人工智能或狭义人工智能
as a weak AI, or narrow AI,a technique
一种擅长或可能会在
that excels,or is maybe even
任务中超过人类
on a superhuman level at a task,
但对别的东西却一无所知的技术
but has zero or no knowledge about anything else.
扩展这些算法的一个关键就是
A key to extend these algorithms would be to design them
将它们的设计成
in a way that their knowledge
能够将知识适用于更多问题的智能
generalizes well to other problems.
我们称之为迁移学习
This is what we call transfer learning,
斯坦福人工智能实验室与
and this collaboration between the Stanford AI
加利福尼亚理工学院间合作进行了一次名为“神经任务编程”项目
lab and Caltech goes by the name Neural Task Programming,
尝试解决这一问题
and tries to tackle this problem.
我们试图解决的任何问题
A solution to practically any problem we’re trying
实际上都可以写成一系列的任务
to solve can be written as series of tasks.
这些典型的复杂动作
These are typically complex actions,
如清洁桌子 或表演一个后空翻
like cleaning a table, or performing a backflip
是很难迁移到别的问题上
that are difficult to transfer to a different problem.
这项技术有点像分步解决型的算法
This technique is a bit like divide and conquer type algorithms
它们积极地将大而困难的任务
that aggressively try to decompose big,
分解成更小 更易于管理的部分
difficult tasks into smaller, more manageable pieces.
越小越易于理解的部分
The smaller and easier to understand the pieces are,
就越容易重复使用
the more reusable they are and the better they generalize.
让我们看一个例子
Let’s have a look at an example.
比如 在需要挑选并放置物体的问题里
For instance, in a problem where we need to pick and place objects,
这个系列任务能够分解成挑选和放置
this series of tasks can be decomposed into picking and placing.
这些可以进一步细分成一系列更小的任务
These can be further diced into a series of even smaller tasks,
比如握紧 移动 以及释放动作
such as gripping, moving, and releasing actions.
总之 如果学习像这样进行
However, if the learning takes place like this,
我们现在可以指定这些任务的不同变化
we can now specify different variations of these tasks,
并且算法很快就能理解
and the algorithm will quickly understand
如何适应这些小任务的结构
how to adapt the structure of these small tasks
来有效地解决新问题
to efficiently to solve new problems.
新算法对具有不同长度 拓扑结构
The new algorithm generalizes really well
和变化目标的任务具有很好的推广性
for tasks with different lengths, topologies, and changing objectives.
如果你看过这篇论文
If you take a look at the paper,
你还会发现更多关于对抗动力学的信息
you’ll also find some more information on adversarial dynamics,
其中列出了一些问题变体
which lists some problem variants
一个可恶的对手不时地
where a really unpleasant adversary pushes things
把东西推到桌面上
around on the table from time to time
来打乱程序
to mess with the program,
并得出一些结论
and there are some results
该算法能很好地从这些失败状态中
that show that the algorithm is able to recover
恢复过来的结论 这真的很酷
from these failure states quite well.Really cool.
现在 请不要把它当成人工智能的完整解决方案
Now, please don’t take this as a complete solution for AGI,
因为这是一件极好的产品
because it is a fantastic piece of work,
但它绝非如此
but it’s definitely not that.
总之 对于最终的解决方案而言
However, it may be a valuable puzzle
它可能是一个有价值的困难部分
piece to build towards the final solution.
这就是研究
This is research.
我们脚踏实地地前进着
We advance one step at a time.
伙计 活着真是太好了
Man, what an amazing time to be alive.
感谢收看 感谢你们的慷慨支持
Thanks for watching and for your generous support,
我们下次见!
and I’ll see you next time!

发表评论

译制信息
视频概述

弱人工智能是啥?又是怎么实现的?人工智能的问题都已经解决了吗?来看看这个关于人工智能和神经任务的两分钟论文吧

听录译者

收集自网络

翻译译者

小葵酱

审核员

审核员MF

视频来源

https://www.youtube.com/watch?v=Lcxz6dtYjI4

相关推荐