Educational Technology: Crash Course Computer Science #39

Summary

这段视频脚本深入探讨了计算机科学如何支持教育技术，从而促进学习。从早期的纸张和铅笔到现代的机器学习智能系统，技术一直是教育的支撑力量。视频提到了教育技术面临的挑战，例如如何为数百万学生提供及时和相关的反馈，以及如何自动化地评分数百万份作业。为了解决这些问题，计算机科学家和教育技术专家正在开发混合的人类-技术系统。视频中还介绍了智能辅导系统，包括生产规则、领域模型和学生模型，以及贝叶斯知识追踪等技术，这些技术可以帮助个性化学习。此外，还讨论了教育数据挖掘如何利用大数据来改进个性化学习。最后，视频展望了教育技术的未来，包括虚拟现实、增强现实以及可能的直接大脑学习技术。

Takeaways

• 🌐 互联网上信息的创造和广泛可用性是计算机技术带来的最戏剧性变化之一，目前有13亿个网站，维基百科有500万篇英文文章。
• 🔍 Google每天提供40亿次搜索，YouTube每分钟有350万次视频观看，400小时新视频被上传。
• 📚 信息的可用性与从中学习是不同的，Crash Course支持互动式课堂学习、指导对话和实践经验作为学习的强大工具。
• 💻 计算机科学可以支持教育技术来促进学习，从纸和笔到基于机器学习的智能系统，技术一直在支持教育。
• 📜 远程教育是教育技术发展的驱动力之一，比如公元50年圣保罗发送的信件，提供了关于宗教教义的课程。
• 📺 教育技术的重大进步，如广播、电视、DVD和激光光盘，都曾承诺要彻底改变教育。
• 🎥 观看教育视频时，可以通过调整播放速度、暂停视频并提问、尝试视频中的例子或练习来显著提高学习效果。
• 📚 MOOC（大规模开放在线课程）是视频传播优质教育的最新形式，2012年被《纽约时报》称为MOOC之年。
• 🤖 教育技术问题吸引了计算机科学家和教育技术专家，他们正在寻找解决方案，如创建混合的人机系统来提供及时和相关的反馈。
• 📈 智能辅导系统（Intelligent Tutoring Systems）使用人工智能技术，通过生产规则（IF-THEN语句）来表示知识、程序和技能。
• 🧠 贝叶斯知识追踪（Bayesian knowledge tracing）是一种常用技术，用于估计学生的知识水平，并根据每次互动的正确性更新估计。
• 📊 教育数据挖掘（Educational Data Mining）利用来自数百万学习者的数据，发现学生常见的陷阱和挫折点，以改善个性化学习。
• 🚀 教育技术的未来可能包括虚拟代理、直接大脑学习等创新，这些技术将提供全新的学习体验。

Q & A

互联网上目前大约有多少个网站？

根据视频脚本，互联网上目前大约有13亿个网站。

Wikipedia有多少篇英文文章？

Wikipedia有大约五百万篇英文文章。

Google每天大约提供多少次搜索服务？

Google每天提供大约四十亿次搜索服务。

在YouTube上，每分钟大约有多少小时的新视频被上传？

在YouTube上，每分钟大约有400小时的新视频被上传。

教育技术在历史上有哪些重要的发展阶段？

教育技术在历史上的重要发展阶段包括从纸张和铅笔的使用，到基于机器学习的智能系统，以及早期的人类洞穴绘画，圣保罗在公元50年左右发送的教义书信，以及后来的广播、电视、DVD和激光视盘等。

托马斯·爱迪生在1913年对教育做出了什么预测？

托马斯·爱迪生在1913年预测，书籍将很快在学校变得过时，电影将能够教授人类知识的每一个分支，并且我们的学校系统将在未来十年内完全改变。

观看教育视频时，有哪些方法可以提高学习效果？

提高观看教育视频学习效果的方法包括：调整视频速度以适应个人节奏，暂停视频并在难点部分进行思考和提问，尝试视频中提供的任何示例或练习，即使不是编程也可以在纸上写伪代码并尝试编码。

什么是大规模开放在线课程（MOOCs）？

大规模开放在线课程（MOOCs）是一种最新的视频传播优质教育的形式，它们通常包括著名教授的讲座视频。《纽约时报》在2012年将MOOCs称为“年度之最”。

当有数百万学习者和少数教学人员时，教育技术面临的主要问题是什么？

当有数百万学习者和少数教学人员时，教育技术面临的主要问题包括如何提供及时和相关的反馈，以及如何评分数百万份作业。

什么是智能辅导系统（Intelligent Tutoring Systems）？

智能辅导系统是一种通常由人工智能驱动的系统，它们使用生产规则和选择算法来代表特定学科的知识、程序和技能，以提供个性化的学习体验，帮助学生在他们最需要练习的地方进行练习。

贝叶斯知识追踪是什么？

贝叶斯知识追踪是一种算法，它将学生的知识视为一组潜在变量，这些变量的真实值对外部观察者（如软件）是不可见的。该算法通过观察学生每次使用该技能的正确性来更新对学生知识状态的估计。

教育数据挖掘（Educational Data Mining）是什么？

教育数据挖掘是一个领域，它允许教师和研究人员从数百万学习者那里收集数据，通过分析学生对问题的回答、他们在提交答案前的暂停时间、视频播放速度的变化，以及他们在讨论论坛上的互动，来发现学生常见的陷阱和挫折点，以帮助改进未来的个性化学习。

未来教育技术可能的发展方向是什么？

未来教育技术可能的发展方向包括使用虚拟现实和增强现实技术提供新的学习体验，以及通过直接的大脑学习，将新技能上传到人的大脑中。科学家们已经在这方面的探索中取得了进展，例如通过大脑信号检测某人是否知道某件事。

Outlines

00:00

🌐 信息时代的教育技术

Carrie Anne在Crash Course Computer Science中介绍了计算机技术如何极大地改变了信息的创造和广泛可用性。目前互联网上有13亿个网站，Wikipedia有五百万篇英文文章，Google每天提供40亿次搜索服务，YouTube每分钟有350万个视频被观看。这些信息可以通过智能手机随时随地访问。然而，信息的可用性并不等同于从中学习。Crash Course支持互动式课堂学习、直接对话和实践经验作为学习的强大工具，同时也相信教育技术在课堂内外的附加力量。视频技术从纸张和铅笔到最新的基于机器学习的智能系统，一直在支持教育。历史上，远程教育一直是教育技术发展的驱动力。教育技术的重大进步，如广播、电视、DVD和激光视盘，都曾承诺要彻底改变教育。尽管如此，视频等格式的教育材料变得越来越流行。观看教育视频时，可以通过调整视频速度、暂停视频思考内容、尝试视频中的示例或练习来显著提高学习和记忆。MOOCs（大规模开放在线课程）是视频传播高质量教育的最新形式，尽管它们并没有像一些人预期的那样取代传统大学。教育技术面临的挑战包括如何为数百万学生提供及时和相关的反馈，以及如何评分数百万份作业。解决这些问题需要创建混合的人类-技术系统。

05:03

🤖 智能辅导系统的工作原理

智能辅导系统，通常由人工智能驱动，可以为学生提供个性化的学习体验。这些系统通过生产规则（IF-THEN语句）来描述解决步骤和学生可能犯的常见错误。生产规则与选择算法相结合，形成领域模型，这是特定学科知识、程序和技能的正式表示。此外，智能辅导系统会构建和维护学生模型，跟踪学生掌握的生产规则和需要练习的地方。为了准确了解学生知道什么和不知道什么，系统使用贝叶斯知识追踪技术，这是一种将学生知识视为一组潜在变量的算法。贝叶斯知识追踪通过观察学生每次使用技能的正确性来更新对其知识状态的估计。教育数据挖掘通过分析学生对问题的回答、他们在视频中的暂停时间、视频播放速度以及他们在讨论论坛上的互动，来发现学生常见的困难和挫折点。这些数据帮助改进未来的个性化学习。

10:08

🚀 教育技术的未来发展

教育技术和设备正在从笔记本电脑和桌面电脑转移到大型桌面表面，让学生可以小组合作，同时也转移到微型移动设备，让学生随时随地学习。虚拟现实和增强现实正在激发新的学习体验，如深海潜水、探索外太空、穿越人体或与他们现实生活中可能从未遇到的文化互动。未来，教育界面可能会完全消失，取而代之的是通过直接的大脑学习，人们可以直接将新技能上传到大脑中。科学家们在这方面已经取得了进展，例如，仅通过大脑信号就能检测出某人是否知道某件事。这引发了一个有趣的问题：如果我们能够将东西下载到大脑中，我们是否也能上传大脑的内容？这一问题将在下周关于计算远未来的系列终章中探讨。

Keywords

💡教育技术

教育技术指的是利用各种技术手段来支持教育和学习的过程。在视频中，教育技术被广泛讨论，包括从纸张和铅笔到基于机器学习的智能系统，这些技术帮助教育实现了跨越千年的发展。视频中特别提到了计算机科学如何通过教育技术支持学习，包括智能辅导系统和个性化学习体验。

💡信息获取

信息获取是指通过技术手段获取和访问信息的过程。视频中提到，目前互联网上有13亿个网站，Wikipedia有五百万篇英文文章，Google每天提供40亿次搜索服务，YouTube每分钟有350万个视频被观看，这些都是信息获取的例子。

💡在线学习

在线学习是一种通过互联网进行的教育方式，允许学生在任何时间和地点学习。视频中提到了MOOCs（大规模开放在线课程）作为在线学习的一种形式，它通过视频讲座等形式提供教育内容。

💡互动学习

互动学习是指学习者通过与学习材料或系统的互动来提高学习效果的方法。视频中强调了互动学习的重要性，如通过暂停视频、提问、尝试视频中的示例或练习等方法来增强学习体验。

💡智能辅导系统

智能辅导系统是一种利用人工智能技术来模拟人类教师的系统，它可以根据学生的学习情况提供个性化的反馈和指导。视频中通过一个假设性的代数问题解决过程来解释智能辅导系统是如何工作的。

💡生成规则

生成规则是智能辅导系统中用来描述解题步骤的一种IF-THEN语句。视频中提到，这些规则不仅可以用来表示正确的解题步骤，还可以表示学生可能犯的常见错误，即所谓的“buggy rules”。

💡领域模型

领域模型是智能辅导系统中用来正式表示特定学科知识、程序和技能的一种模型。它通过生成规则和选择算法的结合来形成，帮助学生解决个别问题。视频中解释了领域模型如何与学生模型相结合，以提供个性化的辅导。

💡贝叶斯知识追踪

贝叶斯知识追踪是一种用于智能辅导系统中的算法，它通过观察学生对每个技能的使用情况来更新对学生知识状态的估计。视频中提到了这种算法如何通过考虑学生已经掌握的技能和通过工作解决问题而学习的概率来工作。

💡教育数据挖掘

教育数据挖掘是从大量学习者的数据中发现模式和趋势的过程。视频中提到，通过分析学生对问题的回答、他们在视频中的暂停时间、视频播放速度以及他们在讨论论坛上的互动，可以改进个性化学习。

💡虚拟现实和增强现实

虚拟现实(VR)和增强现实(AR)是提供沉浸式学习体验的技术，它们允许学生在模拟环境中进行探索和学习。视频中提到，这些技术正在激发新的教育体验，如深海潜水、太空探索和人体内部旅行。

💡直接大脑学习

直接大脑学习是一种假设性的学习方式，它涉及到将新技能直接上传到大脑中，而不是通过传统的学习过程。视频中提出了一个远景，即未来的教育界面可能会完全消失，取而代之的是通过直接大脑学习来获取技能。

Highlights

计算机技术最显著的变化之一是信息的创造和广泛可用性。

互联网上目前有13亿个网站。

维基百科拥有五百万篇英文文章，涵盖从1518年的舞蹈瘟疫到正确的卫生纸卷方向等一切内容。

谷歌每天提供40亿次搜索以访问这些信息。

YouTube上每分钟有350万个视频被观看，用户上传了400小时的新视频。

教育技术，从纸和笔到最近的基于机器学习的智能系统，已经支持了数千年的教育。

远程教育一直是教育技术发展的驱动力。

托马斯·爱迪生在1913年预测，书籍将很快在学校中变得过时。

观看教育视频时，使用视频速度控制调整节奏，确保理解内容并有时间反思。

在视频的难点暂停，问自己问题，并尝试回答，或预测接下来会发生什么。

尝试视频中提出的任何示例或练习，即使是伪代码书写和编码尝试。

大规模开放在线课程（MOOCs）是视频传播优质教育的最新形式。

计算机科学家和教育技术专家正在寻找解决大规模在线学习问题的方法。

有效的学习需要及时和相关的反馈，但如何在只有一名教师的情况下为数百万学习者提供良好的反馈？

使用算法匹配完美的学习伙伴，以及自动化系统进行评分，同时人类完成其余部分。

智能辅导系统通过生产规则和选择算法结合形成领域模型，正式表示特定学科的知识、程序和技能。

贝叶斯知识追踪是一种常用技术，用于估计学生的知识水平，并支持掌握学习。

教育数据挖掘利用数百万学习者的数据，发现学生常见的陷阱和挫折点。

教育技术正从笔记本电脑和台式电脑转移到大型桌面表面和小型移动设备，提供新的学习体验。

虚拟现实和增强现实正在激发新的教育体验，如深海潜水、探索外太空、穿越人体或与他们现实生活中可能永远无法遇到文化互动。

未来，教育界面可能会完全消失，取而代之的是通过直接的大脑学习，人们可以直接将新技能上传到大脑中。

Transcripts

Hi, I’m Carrie Anne, and welcome to Crash Course Computer Science!

One of the most dramatic changes enabled by computing technology has been the creation

and widespread availability of information.

There are currently 1.3 billion websites on the internet.

Wikipedia alone has five million English language articles, spanning everything from the Dancing

Plague of 1518 to proper toilet paper roll orientation.

Every day, Google serves up four billion searches to access this information.

And every minute, 3.5 million videos are viewed on Youtube, and 400 hours of NEW video get

uploaded by users.

Lots of these views are people watching Gangnam Style and Despacito.

But another large percentage could be considered educational, like what you’re doing right now.

This amazing treasure trove of information can be accessed with just a few taps on your

smartphone.

Anywhere, anytime.

But, having information available isn’t the same as learning from it.

To be clear, we here at Crash Course we are big fans of interactive in-class learning,

directed conversations, and hands-on experiences as powerful tools for learning.

But we also believe in the additive power of educational technology both inside and

outside the classroom.

So today we’re going to go a little meta, and talk specifically about how computer science

can support learning with educational technology.

Intro

Technology, from paper and pencil to recent machine-learning-based intelligent systems,

has been supporting education for millennia - even as early as humans drawing cave paintings

to record hunting scenes for posterity.

Teaching people at a distance has long been a driver of educational technology.

For example, around 50 CE, St. Paul was sending epistles that offered lessons on religious

teachings for new churches being set up in Asia.

Since then, several major waves of technological advances have each promised to revolutionize

education, from radio and television, to DVDs and laserdiscs.

In fact, as far back as 1913, Thomas Edison predicted,

“Books will soon be obsolete in the schools…

It is possible to teach every branch of human knowledge with the motion picture.

Our school system will be completely changed in the next ten years.”

Of course, you know that didn’t happen.

But distributing educational materials in formats like video has become more and more popular.

Before we discuss what educational technology research can do for you, there are some simple

things research has shown you can do, while watching an educational video like this one,

to significantly increase what you learn and retain.

First, video is naturally adjustable, so make sure the pacing is right for you, by using

the video speed controls.

On YouTube, you can do that in the right hand corner of the screen.

You should be able to understand the video and have enough time to reflect on the content.

Second, pause!

You learn more if you stop the video at the difficult parts.

When you do, ask yourself questions about what you’ve watched, and see if you can answer.

Or ask yourself questions about what might be coming up next, and then play the video

to see if you’re right.

Third, try any examples or exercises that are presented in the video on your own.

Even if you aren’t a programmer, write pseudocode on paper, and maybe even give coding a try.

Active learning techniques like these have been shown to increase learning by a factor

of ten.

And if you want more information like this - we’ve got a whole course on it here.

The idea of video as a way to spread quality education has appealed to a lot of people

over the last century.

What’s just the latest incarnation of this idea came in the form of Massive Open Online

Courses, or MOOCs.

In fact, the New York Times declared 2012 the Year of the MOOC!

A lot of the early forms were just videos of lectures from famous professors.

But for a while, some people thought this might mean the end of universities as we know them.

Whether you were worried about this idea or excited by it, that future also hasn’t really

come to pass and most of the hype has dissipated.

This is probably mostly because when you try to scale up learning using technology to include

millions of students simultaneously with small numbers of instructional staff - or even none

- you run into a lot of problems.

Fortunately, these problems have intrigued computer scientists and more specifically,

educational technologists, who are finding ways to solve them.

For example, effective learning involves getting timely and relevant feedback – but how do

you give good feedback when you have millions of learners and only one teacher?

For that matter, how does a teacher grade a million assignments?

Solving many of these problems means creating hybrid, human-technology systems.

A useful, but controversial insight, was that students could be a great resource to give

each other feedback.

Unfortunately, they’re often pretty bad at doing so – they’re neither experts

in the subject matter, nor teachers.

However, we can support their efforts with technology.

Like, by using algorithms, we can match perfect learning partners together, out of potentially

millions of groupings.

Also, parts of the grading can be done with automated systems while humans do the rest.

For instance, computer algorithms that grade the writing portions of the SATs have been

found to be just as accurate as humans hired to grade them by hand.

Other algorithms are being developed that provide personalized learning experiences,

much like Netflix’s personalized movie recommendations or Google’s personalized search results.

To achieve this, the software needs to understand what a learner knows and doesn’t know.

With that understanding, the software can present the right material, at the right time,

to give each particular learner practice on the things that are hardest for them, rather

than what they’re already good at.

Such systems – most often powered by Artificial Intelligence – are broadly called

Intelligent Tutoring Systems.

Let’s break down a hypothetical system that follows common conventions.

So, imagine a student is working on this algebra problem in our hypothetical tutoring software.

The correct next step to solve it, is to subtract both sides by 7.

The knowledge required to do this step can be represented by something called a production rule.

These describe procedures as IF-THEN statements.

The pseudo code of a production rule for this step would say if there is a constant on the

same side as the variable, then subtract that constant from both sides.

The cool thing about production rules is that they can also be used to represent common

mistakes a student might make.

These production rules are called “buggy rules”.

For example, instead of subtracting the constant, the student might mistakenly try to subtract

the coefficient.

No can do!

It’s totally possible that multiple competing production rules are triggered after a student

completes a step – it may not be entirely clear what misconception has led to a student’s answer.

So, production rules are combined with an algorithm that selects the most likely one.

That way, the student can be given a helpful piece of feedback.

These production rules, and the selection algorithm, combine to form what’s called

a Domain Model, which is a formal representation of the knowledge, procedures and skills of

a particular discipline - like algebra.

Domain models can be used to assist learners on any individual problem, but they’re insufficient

for helping learners move through a whole curriculum because they don’t track any

progress over time.

For that, intelligent tutoring systems build and maintain a student model – one that

tracks, among other things, what production rules a student has mastered, and where they

still need practice.

This is exactly what we need to properly personalize the tutor.

That doesn’t sound so hard, but it’s actually a big challenge to figure out what a student

knows and doesn’t know based only on their answers to problems.

A common technique for figuring this out is Bayesian knowledge tracing.

The algorithm treats student knowledge as a set of latent variables, which are variables

whose true value is hidden from an outside observer, like our software.

This is also true in the physical world, where a teacher would not know for certain whether

a student knows something completely.

Instead, they might probe that knowledge using a test to see if the student gets the right answer.

Similarly, Bayesian knowledge tracing updates its estimate of the students’ knowledge

by observing the correctness of each interaction using that skill.

To do this, the software maintains four probabilities..

First is the probability that a student has learned how to do a particular skill.

For example, the skill of subtracting constants from both sides of an algebraic equation.

Let’s say our student correctly subtracts both sides by 7.

Because she got the problem correct, we might assume she knows how to do this step.

But there’s also the possibility that the student got it correct by accident, and doesn’t

actually understand how to solve the problem.

This is the probability of guess.

Similarly, if the student gets it wrong, you might assume that she doesn’t know how to

do the step.

But, there’s also the possibility that she knows it, but made a careless error or other slip-up.

This is called the probability of slip.

The last probability that Bayesian knowledge tracing calculates is the probability that

the student started off the problem not knowing how to do the step, but learned how to do

it as a result of working through the problem.

This is called the probability of transit.

These four probabilities are used in a set of equations that update the student model,

keeping a running assessment for each skill the student is supposed to know.

The first equation asks: what’s the probability that the student has learned a particular

skill which takes into account the probability that it was already learned previously and

the probability of transit.

Like a teacher, our estimate of this probability that it was already learned previously

depends on whether we observe a student getting a question correct or incorrect,

and so we have these two equations to pick from.

After we compute the right value, we plug it into our first equation, updating the probability

that a student has learned a particular skill, which then gets stored in their student model.

Although there are other approaches, intelligent tutoring systems often use Bayesian knowledge

tracing to support what’s called mastery learning, where students practice skills,

until they’re deeply understood.

To do this most efficiently, the software selects the best problems to present to the

student to achieve mastery, what’s called adaptive sequencing, which is one form of

personalization.

But, our example is still just dealing with data from one student.

Internet-connected educational apps or sites now allow teachers and researchers the ability

to collect data from millions of learners.

From that data, we can discover things like common pitfalls and where students get frustrated.

Beyond student responses to questions, this can be done by looking at how long they pause

before entering an answer, where they speed up a video, and how they interact with other

students on discussion forums.

This field is called Educational Data Mining, and it has the ability to use all those facepalms

and “ah ha” moments to help improve personalized learning in the future.

Speaking of the future, educational technologists have often drawn inspiration for their innovations

from science fiction.

In particular, many researchers were inspired by the future envisioned in the book

"The Diamond Age" by Neal Stephenson.

It describes a young girl who learns from a book that has a set of virtual agents who

interact with her in natural language acting as coaches, teachers, and mentors who grow

and change with her as she grows up.

They can detect what she knows and how’s she’s feeling, and give just the right feedback

and support to help her learn.

Today, there are non-science-fiction researchers, such as Justine Cassell, crafting pedagogical

virtual agents that can “exhibit the verbal and bodily behaviors found in conversation

among humans, and in doing so, build trust, rapport and even friendship with their human students."

Maybe Crash Course in 2040 will have a little John Green A.I. that lives on your iPhone 30.

Educational technology and devices are now moving off of laptop and desktop computers,

and onto huge tabletop surfaces, where students can collaborate in groups, and also tiny mobile

devices, where students can learn on the go.

Virtual reality and augmented reality are also getting people excited and enabling new

educational experiences for learners – diving deep under the oceans, exploring outer space,

traveling through the human body, or interacting with cultures they might never encounter in

their real lives.

If we look far into the future, educational interfaces might disappear entirely, and instead

happen through direct brain learning, where people can be uploaded with new skills, directly

into their brains.

This might seem really far fetched, but scientists are making inroads already - such as detecting

whether someone knows something just from their brain signals.

That leads to an interesting question: if we can download things INTO our brains, could

we also upload the contents of our brains?

We’ll explore that in our series finale next week about the far future of computing.

I'll see you then.