Play then build: a learning design pattern
Decades ago, in 1997, IBM supercomputer Deep Blue managed to edge out chess world champion Garry Kasparov in a tournament that has gone down in history. These days, garden variety chess programs — able to run on a typical smart phone — can beat even the strongest human players.
And yet computers are only now getting to the point of being somewhat adequate at detecting whether or not a photo has a cat in it.
The world of artificial intelligence is complicated, to say the least.
Here’s a recent challenge I had as a learning designer:
We had an existing course that comprehensively covered the field of AI for computer science students. We wanted to use this as a basis to create a short course for a target audience that might not be so technically literate. There’s a lot of interest and intrigue in AI right now, and the goal of this short course was to demystify the whole venture, particularly machine learning.
If artificial intelligence is complex, machine learning borders on the mystical.
So, on the one hand, we have an eager but unfamiliar target audience. On the other, we have a tricky subject matter. This called for just the right learning design! I used a particular kind of scaffolding that I call: Play then build.
It consists of three main steps. Let me explain each step as if it’s in a movie.
At the movies: Lights, camera, action!
You’ve just found your seat and you’ve barely started eating your popcorn when all of a sudden, bam!, you’re in the middle of a getaway chase scene. It seems the bad guys have just staged the world’s biggest bank heist, and every last police car in the city now appears to be hot on their tail.
And what’s this… Brad Pitt, who you were sure is playing the hero FBI agent in the movie, is sitting in the back of the getaway car, pulling off his balaclava!
You’re not entirely sure what is going on. You feel as though you’re watching the film’s climactic third act at the beginning for some reason. But you also know this is why you paid for the tickets. This is what you came to see.
Step 1: Play with the goal
My first step was to encourage the learner to get a feel for what machine learning is by discovering what machine learning can do. I leveraged my target audience’s intrigue in the subject and said: Here, play with this!
For example, did you know that you can train a program to learn different visual cues (which you provide via your webcam) and to then play Pac-Man based on those visual cues? Fun, right?
Obviously, there’s a lot going on under the hood here, but why start with that? Why not jump to the climax, to see where this is all going?
I gave the learner the opportunity to explore applications like these, because:
- They are fun and interesting, and what the student came to see
- They are practical and they showcase how this technology can be applied
- They give the student a sense of how machine learning works — and why older/traditional algorithm technologies won’t let AI do all that we want it to do — even if they don’t understand everything that’s going on (yet)
Tips for Step 1
Here are some tips for an effective Play opening.
1. Don’t just show the crazy explosions without context.
The idea is to put the interesting applications front and centre. To showcase “what you can do with all of this”. You might be tempted to turn the opening into a line up of the coolest applications you can find. But the learning strategy here is still to give the learner some bearings and some context.
For example, in my course, when introducing Webcam Pac-Man, the learner was encouraged to consider what sort of machine learning was going on here. What task(s) was it learning to perform? What data was it learning from?
2. Let them play.
The obvious difference between this approach and a more traditional textbook approach is that the learner starts with the “end”. Instead of starting with “here is the history and foundation of this topic”, the learner starts with “here’s what you can do with it”.
The operative word here is “do”. This isn’t just a difference in the order of when things are explored. It’s a difference in how they are explored.
Don’t just tell them, don’t just show them. Work out how to enable the learner to get their hands dirty. For example, Webcam Pac-Man lets you experiment with different inputs, and explore the effectiveness of the learning. (It also lets you actually play Pac-Man, so there’s that.)
3. Curiosity is good!
While the learner is getting context, they won’t understand everything yet. That’s OK. If you’ve designed the Play right, they should be curious about how the game underneath works. To help ensure that, you should at least build some implicit (unanswered) questions into your Play.
Or, better yet, why not make them explicit? I asked my learners to consider: Think about how Webcam Pac-Man is performing the task it is performing. What makes this learning? Why won’t “traditional” algorithms work here?
At the movies: Two weeks earlier…
Brad Pitt pulls his balaclava off, and you gasp with the rest of the audience. How could this be? Isn’t he the hero? Why is he with the crooks?
All of a sudden, the screen cuts to black and the words appear:
TWO WEEKS EARLIER…
And we fade in to a less frenetic scene, where Brad Pitt is now wearing typical detective attire. A subtitle tells us we are at the FBI headquarters. Brad’s on a case. All is good with the world. In fact, he is working undercover to infiltrate a bank heist gang. That must be why he ends up with them at the end!
For now, let’s sit back and find out who these guys are and what they’re up to.
Step 2: Explore the simplest case possible
Once I gave the learners the opportunity to play with the goal, I took it back to basics, back to the beginning. The second step involved going over an incredibly simple application of machine learning. Specifically, we looked at:
How can a program learn to detect whether an email is spam?
This was my way of encouraging the learner to explore the foundations of machine learning. The “moving parts” are relatively simple to follow:
- With a bit of exploring, it’s relatively easy for the student to see why machine learning is used in this situation, and why a traditional algorithm wouldn’t be able to keep up with such a task
- Spam filtering is also a simple enough use case to allow us to explore “what’s under the hood”, i.e., what the main components are.
So the learner was able learn “how machine learning works” in a robust yet simple environment, even if it isn’t the most cutting edge technology!
Tips for Step 2
1. Make it clear you’re keeping it simple.
There’s a reason those movies that use the above “flashback” trick need to say “TWO WEEKS EARLIER”. If they didn’t, the audience would lose its bearings. Similarly, you need to make it clear that, during this part of the course, we’re going back to basics.
In the spam filtering example, the learner was aware that this wasn’t the most cutting edge technology, and the application of machine learning wasn’t quite up to the examples they were just using during Play. But there is a sense of reassurance here: we are taking our time to lay the foundation of knowledge with a simple, universally understood, example.
2. Use a prototype.
Developing an appropriate example in this part of the course may be the most intricate part of your design. You need to give the learner a “prototype” that has as many of the moving parts as possible, without being overbearing or bogged down in detail.
At the movies: How did they do that?
The first act of the movie played out fairly calmly. You got to meet the protagonist and understand their world.
But this isn’t just a movie about an FBI detective. It’s a movie about a bank heist investigation, and how that detective got caught up in it. Because you already know that he does, the only question now is how…
Step 3: Build the complexity
By now, the learner has played with the goal, and through that they’ve gotten some context about where all of this is heading: they’ve seen the forest. They’ve also seen how to plant a tree, by exploring the foundations of the subject matter with as simple a prototype as possible.
Now, it’s time to explore some other trees in these woods. The final step — and this may make up the majority of the learning experience — involves building up the understanding.
In my case, we explored more and more complex uses for machine learning. How can opposing programs learn how to play hide-and-seek from each other? How does a robotic device learn to move around its environment?
Heck, how can a computer learn to detect whether a photo has a cat in it?
There are a lot of moving parts to machine learning that “spam filters” just won’t address. My guiding force in Step 3 was to design use cases so the learner can explore those moving parts, all the while validating the big picture that they glimpsed at the start of the learning experience.
Tips for Step 3
1. Fill the gaps.
Just as your “prototype” example needs to be carefully designed, selected, and finessed, so to do all of your use cases and applications in Step 3. Specifically, each different application you bring in should have a clear gap being filled.
For example, spam filtering allowed the learner to explore what’s known as supervised learning to classify objects made up of words (i.e. emails are either spam or not, depending on the words they contain). Bringing in pictures (i.e. pictures of cats!) brings in a whole new dimension of complexity, so that was the motivation for that use case.
2. Validate the goal.
In Step 1 you gave the learner a taste for the big picture. Step 3 is all about validating, and enabling the learner to validate, that big picture. Use checkpoints to encourage the learner to orientate themselves by that big picture. How close are they getting to it? What’s missing? What do I need to learn and do that complete my journey?
Where to from here?
If you want to give this pattern a go, first ask yourself:
Is the pattern right for your learning needs? I would typically recommend this approach when (1) Your subject matter is either quite technical or in some way niche and hard to “crack”; and (2) Your target audience is general enough that you can’t assume they have a high level of “niche literacy”.
Assuming the learning pattern is right for you, remember these steps:
Step 1: Play with the goal. Showcase why such a niche or technical subject matter is worth exploring. Show the learner what they can do with it.
Step 2: Explore the simplest case possible. Take it back to the “beginning”. That doesn’t mean giving a history lecture — it means enabling the learner to explore the basic moving parts.
Step 3: Build the complexity. Using Step 1 as your North Star, build on Step 2 and let the learner explore the field in more and more complexity.
Have you ever used a pattern like the one I’ve described, or a variation? Tell me all about it by commenting below!
