AI has long been treated with disdain by the education community because of its potential to replace teachers with robots. This rhetoric is widely spread even while the country faces one of the most acute K–12 teacher shortages and largest inequalities in achievement in the history of time. Instead of a world where teachers are replaced with AI-powered intelligent tutor systems (ITSs) and MOOCs, Varun Arora presents an alternative future, one where teachers are the solution and AI is an important tool in their arsenal.
Varun shares some of the recent successes in deep learning, reinforcement learning, and knowledge representation and extraction in the context of the most complex teaching and learning challenges in K–12 and higher education today. Varun also dives into some of the more unique challenges faced when building models in the educational domain, particularly as they relate to significant training data, privacy, and lowering bias, and highlights applications of pattern recognition AI solutions in classrooms today, along with data on their effectiveness. Varun concludes by exploring open challenges in the teaching and learning domains, the state of current research, and areas for most impact in education and economy upskilling in America today.
4. Obligatory slides on AI being more than supervised machine learning
Knowledge
Representa
tion
Robotics,
Motion,
Planning
Natural
Language
Processing
Social
Intelligence
Speech
Computer
Vision
Expert
Systems
Pattern
Recognitio
n
Machine
Learning
5. The focus today
▪ Scientific learning work in bubble, somewhat agnostic to educational climate
▪ Cognitive Mastery systems:
▪ Intelligent Tutoring Systems (ITSs)
▪ Intelligent Testing Tools
▪ Beginning to see growth of Design-Based Implementation Research
6. Looking beyond intelligence in student screen-time
▪ “Once students bring the device,
artificially intelligent teaching and learning technology development will follow”
▪ 15 years ago: too expensive
▪ 7 years ago: high total cost of ownership
▪ Today: (fill new excuse here)
7. Shift focus to perception
through vision and speech
“AI as a TA” - Rose Luckin
8. 7 key areas of opportunities for AI
▪ Feedback and scoring
▪ Improving college readiness
▪ Empowering students with physical and learning challenges
▪ Behavior Management + Social and Emotional Learning
▪ Curriculum development and alignment
▪ Deeper, higher-order and authentic learning
▪ Teacher ed and in-service professional development
▪ Feedback and scoring
▪
▪ Empowering students with physical and learning challenges
▪
▪ Curriculum development and alignment
▪
▪
10. Feedback and Scoring: Problem Space
▪ Critical mechanism to human learning
▪ Essential to high and low-stakes assessment
▪ Teachers bogged down by piles of grading
11. Feedback and Scoring: Problem Space
Structure query
Seek / get input
Perceive input
Understand input
Compare with expected response
(Use rubric to classify quality)
Use delta from comparison to construct feedback
12. Feedback and Scoring: Problem Space
Structure query ▪ Action (either as a gesture or as performance)
▪ Selection from choices
▪ Boolean response
▪ Single word / phrase
▪ Open-ended short written response
▪ Open-ended long written response
13. Feedback and Scoring: Problem Space
Seek / get input
Perceive input
▪ Written response
▪ Oral / speech response
▪ Drawing
▪ Affective states
▪ Gestures
▪ Computer inputted
14. Feedback and Scoring: Problem Space
Structure query
Seek / get input
Perceive input
Understand input
Compare with expected response
(Use rubric to classify quality)
Use delta from comparison to construct feedback
17. Feedback and Scoring: Opportunity Space
Understand input
Compare with expected response
Use delta from comparison to construct feedback
18. Empowering students w/ physical + learning challenges: Problem Space
▪ Unfair learning opportunities for children with ADHD, Autism Spectrum Disorders (ASD) incl. dyslexia,
dysgraphia, speech disorders, etc.
▪ Teachers have to make accommodations in every instructional activity
▪ Special Ed support is hard for schools to provision
▪ Even families struggle with providing the right support at home
▪ Children inexpressively are hoping to understand and to be understood
19. Empowering students w/ physical + learning challenges: Opportunity
Space
▪ Multimodal learning from AVSR
▪ Simultaneous machine translation
▪ Spectrogram adaptation
▪ Conversational agent modeling
Hoping to understand Hoping to be understood
▪ Understanding facial expression, affective
states and emotion
▪ Perceiving actions
▪ Generative modeling:
▪ Speech synthesis and voice cloning
▪ Using natural language generation to
construct responses
20. Curriculum development and alignment: Problem Space
▪ Curriculum doesn’t work for most students today; personalized learning not necessarily digital
▪ When curriculum works:
▪ Students are mostly engaged.
▪ There would be no times when students would feel entirely lost / confused, unless very intentionally placed in
those states
▪ They would continuously and equitably reach age-level expectations
▪ Curriculum development and alignment is seriously non-trivial
21. Curriculum development and alignment: Problem Space
Knowing exactly
where most of
your students
are
Knowing details
of competency
expectations +
pacing
predictions
Pedagogical
Content
Knowledge and
sequencing
insights
Knowing how
learning will be
assessed
Knowing how
your students
are feeling
Knowing how to
and constructing
materials
Knowing
practical
resource
limitations
Knowing the
content
22. Curriculum development and alignment: Opportunity Space
Knowing exactly
where most of
your students
are
Knowing details
of competency
expectations +
pacing
predictions
Pedagogical
Content
Knowledge and
sequencing
insights
Knowing how
learning will be
assessed
Knowing how
your students
are feeling
Knowing how to
and constructing
materials
Knowing
practical
resource
limitations
Knowing the
content
23. Where do we go from here?
▪ Data
▪ Accuracy
▪ Need for domain-specific research and
practice
Challenges
Need for an Open AI for
education
Be grateful for presence
Not going to waste time on my intro, the session info has it
Get a survey of room distribution, will help with how much technical depth need to cover, since this is technically a business talk
Before starting to talk about teachers, students, and AI
Fear sounding repetitive and meta, but important to step back on the many things AI is
Everyone knows that its a field dedicated to make machines do the cognitive functions that humans do
That sounds extremely general and viscerally omnipresent (search to find things, databases to store knowledge, any sort of interface personalization like any feed, perhaps everytime you call a hotline number and they try to guide you through their maze - as if the system understands you)
I call this idea the AI-presence paradox; and ending up on the realization side of this paradox helps us understand that we have been an AI journey for several decades, and instead of resisting “AI”, we shape AI to be useful to us.
Yes, what we don’t have is general intelligence.
Particularly important in education where are fearing robot teachers
It also turns out that under this “broader” definition.- domains that are constrained by extreme pressure on scaling human capital for well-understood tasks are in the best position to take advantage of handing off laborious perception and cognitive functions to machines
If AI been around for so long, why are we suddenly talking about it a lot more now and throwing conferences on it
We have made some master successes on supervised machine learning - specifically neural networks - in the past decade or so, and so it seems like AI = Supervised ML.
For the unacquainted, machine learning usually entails a computer determining some kind of way to make a prediction. Like a strong approximation function. So its pretty statistical in nature.
Like pointing a webcam or a phone camera to the sky and saying, hey computer camera, you see the sky right? Can you guess what the weather is like?
Now I think even humans suck at that. Actually humans suck at all tasks we have not done before. Only once we have done something several times, we learn the right patterns of doing things.
The bit about “doing something several times” - gives our minds the necessary data. When we make predictions based on data, that’s when we actually call is supervised learning.
So for a machine it would be giving it lots of pictures of the sky, and telling it, “hey, for this day the weather was 102 degrees” and so on
The second idea that is important to understand is that not only is AI not a discipline limited to math and computer science, it characteristically involves more disciplines like psychology (particularly childhood development and human learning), sociology, neuroscience, linguistics also.
This is often overlooked because of the job market around AI
Even within the narrower context of anything remotely computer related, it has a large number of areas that are not supervised machine learning
Why do I care to point these generic things out in a talk on teaching and learning?
Because it turns out that a large number of educators and learning designers and scientists have been trying to do work in AI for quite some now without attributing or announcing it as “AI”
As a “textbook” example, you know at all educational institutions, we record into some spreadsheet course or unit learning outcomes to lead up to some broad list of skills that we think is constitute the progress of our students in becoming good in a topic area.
AI because we are trying to model the student thinking outside their brain and then build technological and non-technological support structures to trick their minds to acquiring these at an expedited pace. It’s also a prerequisite to a computer assisting you with instructional design
Once we start acknowledging it as AI, we can suddenly use a gush of new literature and tools that weren’t at our disposal
[4 mins]
Let’s talk about the case when people have been calling it AI.
Sadly, we have been seeming to not be as inclusive as we could be.
Part of the reason is because for every one person who puts the AI as a support to the teacher, there are 5 clueless idiots promoting the rhetoric of robots replacing teachers.
What are all those people - putting AI in support of the teacher - doing though?
They are doing scientific research in making the learning process better and better - either by advancing on the learning theory OR adapting newer techniques from machine learning. Not always holistic in an education setting sense (I’ll explain what I mean by not holistic in an education setting). This is both good and somewhat bad.
We have made a lot of progress either in learning science in offline learning and Massive Online courseware OR in cognitive mastery systems
Cognitive mastery basically means doing becoming proficient mentally at something (there are tens of consumer apps that make you obsessively repeat a process making you good at it)
ITSs are a particular breed where the goal is to master some subject material - of various scope - such that you keep advancing to more difficult materials. The computer keeps a track of what you have learned and by feeds you the next material by making prediction of what is the best pathway. Its not “intelligent” if it just has a preset list of questions
Testing tools are geared towards taking your inputs and giving feedback, but not so much on teaching you through scaffolded exercises
But we rarely work with constraints and practical limitations within real classrooms and in non-pilot settings. These constraints include hardware deployments, local teacher (or faculty, if we are considering higher ed) professional development cycles, and most importantly scaling beyond one pilot unit and 200 questions, and 2 machine learning models. Often financing is an issue.
Some of these pains hope to be resolved with DBIR, but they are not necessarily making bringing more AI tools scale
DBIRs involve carefully designed studies with school systems as stakeholders and designers from day one, to take into their classrooms
Now if we think the kinds of things we are doing is good enough, the question we can ask is “how are we doing?”
We are certainly making progress, but very humble. We have a pretty decent foundation.
We are far slower than the pace at which AI research / productization is happening. We have tools and opportunities to expedite our investment, unlike during AI winters.
Now I have a strong opinion on this topic with my time in classrooms and working in AI.
A large amount of the work in bringing “intelligence” to classrooms starts with the assumption that it can only happen when every student has a laptop or phone on them.
Sometimes making it feel like it is the only thing preventing every kid’s life changing
Now 15 years ago the excuse was computer penetration: laptops and cellphones are too expensive, make them cheap, and then we will build great adaptive learning experiences
7 years ago we said: computing has become cheaper, but hey, total cost of ownership is the reason schools can’t have computers in everyone’s hands
Today, one thing is clear: cost of the device is not the problem, nor has the world made magical successes in adaptive learning.
Cell phones are dirt cheap. Put aside rural K-12 schools in Ghana; almost every student at private universities in the US has at least one computing device that they bring to the classroom, tuition rates are through the roof. Even then, “intelligent" tutoring is pretty non-existent
Its in a chicken-and-egg limbo; a lot of entrepreneurs and researchers around the world have tried, but they complain about school unwillingness to want their intelligent systems, and schools will consistently complain about these not meeting their needs or not being good enough. So much for intelligence and personalization
I think that by limiting our view of bringing intelligence to the classroom - primarily through student screen interaction - we are making it difficult for teachers to experience intelligence because teacher-role change requires a massive paradigm shift in how we perceive schooling, which makes me believe that it is not coming in the coming decade
It is also true that until we have a value proposition so high - which is a much more advanced state and quality of ITSs, there is going to be apprehension and resistance
[8 mins]
So what is low resistance and high value where I think a bulk of our work should focus on in the near future: using our successes in perception to make the teachers more powerful in physical classroom environments.
Use what students naturally show and say to give an AI necessary inputs to figure out how to guide the teacher in pedagogy and curriculum.
AI as a TA (needs to do the cognitive heavylifting). Which Rose Luckin came up which perfectly summarizes my view on how AI should be positioned within the classroom.
Shift in focus possible due to recent approaches to speech and vision
This actually changes the dynamic significantly of AI in education because being inside more classrooms becomes a whole lot easier, obviously conditional to and thanks to some major machine learning advancements, which I’ll talk about shortly.
I have identified 7 areas where recent successes in AI can give us the greatest returns in teaching and learning
Will be in soon-to-be-published handbook
will focus on 3 (because of the time limit) because they are uniquely important in their own ways
For each area, I’ll give you some state of the problems in simple words that most people can understand. And I’ll give you some ideas for where AI can help - by fitting most naturally. They are all based in this AI as teacher’s TA framework.
Very high-level overview / primer - can’t get into any technical depth or focus on specific papers or techniques in short time - also because this a part of the AI Business Summit.
Because this is a thing we don’t think about it everyday, its important that we lay a common language. Each of these are several hundred sub-problems, and for each subproblem, there are sub optimizations and sub techniques. Ofcourse there are many many companies also trying to solve parts of these challenges
Also, if you are in a role where investment can be made in AI to improve teaching and learning, this is a little help to structure your thoughts
[11 mins]
In some ways represents the entire spectrum of AI problem solving - you can move along a continuum of shallow AI all the way to AI completeness
Stating the obvious
Extremely useful in diagnostic assessment, and formative assessments. [Do a formative assessment by asking what formative assessment is]
Teachers are bogged down by piles of grading, and is one of the biggest reasons for frustration in the industry - no time.
Thats at a time when most students get too little too late feedback. So the gap is huge.
Lectures are the best resort when you don’t know of a problem - think about if someone comes to you for advice, if you don’t gauge specific problems they are facing, you go into lecture mode
Feedback processes look like (NOT including character / engagement feedback):
compare with gold standard is a learning stepLast step: you can also offer supporting resources.
Humans do this fairly fast (also makes it ripe as it passes Andrew Ng’s <1 second litmus test) when shallow thinking is involved (go through an example?)
And all of these in what most technical people would call sequences - science lab experiment (walk through example)
All performed in a variety of settings
(in independent work - in class or at home, whole-group setting, small-group, one-to-one, peer interaction).
Note that corresponding to each input, there is a latent cognitive process that is critical but mostly fail to measure (the exception being “show your work” kind of problem solving)
Understanding it means codifying / labelling it. 4 possible student solutions - short phrase responses.
Taking student solutions and classifying them (on paper). This is a simple example to illustrate a much bigger idea that unlabelled data or the lack of ability to label is generally much less useful in AI
Rubrics useful for complex response types: because associating a single number / grade is complex for complex response types is hard for humans. In simple response type, your obviously just looking to find a match
Delta from the rubrics will give you efficiently what was wasn’t right - but it won’t tell you was wrong. BIG DIFFERENCE.; remedial practice for solving the problem
We are going to come back to structure query in curriculum alignment and development
Input has to be produced by students; there is nothing anyone would ever want to change about that. sure, we can engineer better query or knowledge distribution tools that personalize or do differentiation
Handwriting: A series of breakthroughs in deep learning attributed to an idea called convolutional neural networks (apart from a few more nifty tricks like LSTMs) have allowed us to get drastically high accuracy that the OCR systems you used as teenagers. It begins with drawing bounding boxes around words on the document. Complimented by language models - smarter dictionaries. Single shot detection gives techniques to not have a corpus for former text we have trained on.
Moving away from assumption of perfectly scanned text: real world - perspective distortion - computer vision
Speech: ASR through spectrograms or MFCC, also through recurrent neural networks - using models trained “end-to-end”. Speaker identification. Far-field audio recognition and speaker diarization.
Classroom settings: noise removal
- Actions: gesture recognition and other techniques in computer vision to track motion, again through DNNs
Understanding: NLU (go into parts of speech, statistical techniques to understand structure, extract entities…, knowledge components / schemas), understanding sequences
COMPARE WITH GOLD STANDARD: What learning and prediction in a supervised setting means / looks like - EXAMPLE
Construct statements of support: also a learning step many times because highest rating sentence / groups of sentences. NLG. Determine things to include, and then needs to convey the meaning most efficiently - the tools involved in this process (labelling and post) test against syntax and morphology
We are good at individual things in ideal conditions and “well annotated” learning environments and curriculum. Bad at sequencing in generalizable worlds
[21 mins]
Hoping to understand AND hoping to be understood
[25 mins]
My favorite area and probably the area where education will bring a whole lot of novel constraints and opportunities in AI
There is a lot of non-education people over my life who have asked me for evidence. In some ways I am amazed at what is so obvious.
Its not rocket science
A rather simple way to tell that this is true is that most teachers I know who have been teaching for more than a decade fear formally engineering writing curriculum. If you are lucky, teachers are winging “adaptation” of the textbook, otherwise its textbook-ing to a pre-determined pace all the way.
It’s not a linear process, but some aspects of this process include:
Knowing exactly where most of your students are (including in the middle of teaching the curriculum), and what they already know.
Knowing details of competency expectations for the grade AND for the particular unit or topic currently being taught, AND constantly making predictions of how much time to spend on what instructional activities (this is alignment, this is critical because learning of learning is not fully compatible with the everyone-moving-forward together perspective in education)
Competency expectations try to enumerate levels of fluency and transfer, which is not merely choosing topic names
Knowing what kinds of instructional techniques work for what kinds of competency expectations (pedagogical content knowledge), and how best to sequence them
Knowing how learning will be assessed
Knowing how your students are feeling based on time of day, day in instructional session, and in life beyond the classroom
Knowing how to and then constructing materials that will aid in meeting the competencies
Knowing practical resource limitations
Knowing the content (content knowledge)
The best solutions in this space will bring together these seemingly disconnected elements in a way that doesn’t feel less robotic, and more expert assisted
Knowing exactly where most of your students are (including in the middle of teaching the curriculum), and what they already know.
(we talked about in feedback)
Knowledge tracing (bayesian and deep)
Knowing details of competency expectations (or knowledge components) for the grade
NLU and encode them through multi-dimensional embeddings
The past two decades were significant human work to determine these and “codify” them
Reconciliation with instructional activities is a “learnable” problem through feedback/assessment loop, because humans suck at this
constantly making predictions of how much time to spend on what instructional activities
Also learnable problem, with a human-in-the-loop to train right
Knowing what kinds of instructional techniques work for what kinds of competency expectations (pedagogical content knowledge), and how best to sequence them. When this is done in the case of ITS, its sometimes called adaptive content selection
Learnable problem, but a good initializer labelling would be from master teacher labelled examples
Knowing how learning will be assessed
Combination of OCR on corpus of assessment banks, and NLU with ability to predict competency expectations / instructional activity associations
There are ways to inject healthy bias when assessments don’t reflect our desired forms of assessment; especially in rote exam obsessed environments
Knowing how your students are feeling based on time of day, day in instructional session, and in life beyond the classroom
Affect recognition AND just capturing any data on neurological states through conversations/surveys - also facilitated by AI
Knowing how to and then constructing materials that will aid in meeting the competencies
Generative modeling
Materials mean a lot of things: activities, handouts, supporting texts and media (including diagrams),
Begins with extracting and parsing knowledge from a textbook
Natural language generation - STATE OF ART - producing questions based on entities, relationships, and ability to confuse. Also extensive work in generating summaries.
Even creative texts are being authored by machine learning
But that doesn’t lead to any complex learning - it leads to isolated individual questions. RL is a field of a huge amount of promise where in the target is discount sum of long-term rewards. Robot example. The same techniques that help robots figure out long-term thinking, allow us to
Knowing practical resource limitations
No need of AI here; the software doing the work just needs access to the school’s bank account and hopefully not the teachers.
Knowing the content (content knowledge)
We have talked about knowledge retrieval. But here is where machines will be weak for a while before we make some cognitive breakthroughs. Because we can easily respond to facts, but AI cannot learn bigger ideas just yet
[35 mins]
Your like “cool story dude”, so where do we go from here? How do we move away from your theoretical abstract pointers and bring a lot of these things to life? And where should we focus?
OR why haven’t we made some solutions in the classroom already?
challenges:
Data
Accuracy
Need for domain-specific research and practice
Makes it very very difficult for entrepreneurs or even pre-established education companies (nearly all doing some form or the other of AI, example ETS doing a lot of stuff around language and speech) to get very far.
So what is my core suggestion: we build AI infrastructure collectively for teaching and learning. Like an Open AI but for education. If you have the capacity to implement techniques from other industries with your smaller datasets of subsets of these problems, all power to you. Actually, that’s what you could do. Treat this segmentation of problems and opportunities as a template and attack one-problem after another (not necessarily the low-hanging fruit) for the problems you care about most and how deep your pockets are.