Blog post —
Deep Learning: The Machine has no Yet
by Dr. Claudia Rademaker, Co-founder Dugga
Deep Learning
Two words that mean opposite things, depending on who is in the room.
The machine has no "yet"
Two words keep turning up on both sides of my working life: deep learning.
In education, deep learning means a student who genuinely understands something, not one who memorizes it on Tuesday night and has lost it by the following week. Different systems have their own name for it: deep understanding, mastery, conceptual learning. They all point at knowledge that still works when the question is posed differently and over time.
In artificial intelligence, we use exactly the same two words, but we mean the machine. The machine is getting deeper.
Same words, two quite different processes. Moreover, in a few ways they can be opposite ones. It's worth a minute to look at how.
Three differences
1. A machine gets deeper by adding. A person gets deeper by taking away.
Depth in a neural system network is literal: more layers, more parameters. The better system is the bigger one. However, human depth runs the other way. A student who understands proportion is holding fewer things than one who does not. They can carry a hundred worked problems compressed into a single idea anywhere. The machine accumulates. The person distils.
2. A machine needs millions of examples. A person may need one.
Show a model a cat ten million times and it learns cats. However, show a two-year-old a cat just once and that's often it: cats from behind, cats in cartoons, a cat drawn badly in crayon. That is not a difference in efficiency. There’s a different thing happening and it's why a handful of well-chosen problems can do more than another entire chapter of content.
3. A machine cannot be confused.
This is the one I keep coming back to. A model has no experience of not-knowing. Ask it something past the edge of what it learned and it doesn't pause or feel around for footing. It answers wrong with exactly the fluency it has when it's right. It has no yet.
A student does. That particular discomfort. For example; I follow this up to here, and then I lose it. This is not a failure of learning. It's the instrument. A student who can find the edge of their own understanding can aim at it. Every fourteen-year-old has that instrument. No model has it. It is the whole engine of deep learning.
Which draws a useful line. Machines are very good at the work around understanding: building the test, marking it, tracking it, spotting the twelve students who all missed the same step. The understanding itself they cannot do and they cannot tell so. They will report progress either way.
So which one is getting deeper?
That leaves a school with one question to carry into any decision about AI: which one is getting deeper here; the machine, or the student?
For us the order is clear. The machine takes the work that eats the teacher's time. The teacher spends that time on the student. The student goes deeper. If it ever runs the other way around, we've built the wrong thing, whatever the dashboard says.
It's a real constraint on us, and a welcome one. Technology can carry the load around learning, the volume, the repetition, the administration. Heavy work but not specific pedagogical work. This is where a teacher’s pedagogical expertise comes in. For example; what this student actually needs, what they need next, when to push and when to wait. This is not a gap waiting for a better model, it is a different kind of knowing, built from experience in classrooms and from knowing this particular student, in this room, this term.
As such, a teacher's expertise is not a fallback for what modern technology including AI has not reached yet. It's what the technology is there to protect and make room for. That's where we start when we design anything at Dugga: what does this free the teacher to do, where can we enhance teachers’ expertise and does the judgement stay where it belongs? It seems a fair question to ask of any tool a school brings in. Ours included.
What a test tells a student
Every test a student takes tells that student something about themselves. That they are capable, or that they are not. And most students remember that message far longer than they remember the subject it was attached to.
So when assessment goes badly, the real damage is not a wrong grade. It's a fifteen-year-old who has quietly concluded they are not a mathematics person. That conclusion tends to last a lifetime, and it is very often simply untrue.
When it goes well, something else happens. The student finds out precisely what they don't understand yet - and that word, yet, is the whole thing. Not a verdict. A description of where they're standing today.
Students assessed that way are less afraid. They ask more questions. They tell a teacher when they're lost, because being lost has stopped being an admission of failure. And a student who hides confusion is invisible to us, while a student who shows it can be helped.
The machine has the same blind spot, for the opposite reason. A model cannot tell you where its understanding runs out. A frightened student can: they know the exact lesson where they lost the thread but won't say it out loud. Either way, nobody in the room can see where understanding ends. Only one of the two is fixable, and that's what good assessment is for: not grading the confusion but making it safe to name. Once it's named, it takes a teacher to know what to do about it.
The instruction sheet
Assessment is not the last step in education. It's the instruction sheet. Whatever gets measured is what ends up being taught, which makes it hard to ask for real understanding while still testing for recall. The classroom follows the exam, not the curriculum document.
Currently, many education systems are making two big bets at once: writing deep understanding into their curricula and bringing in AI faster than anyone can properly evaluate. Both get settled in the same place, not in the policy document, not in the procurement meeting, but in what a student is actually asked to do on the day of the test.
Worth getting right, then. A machine will never need the word yet. Every student does.
That's the work we continue to do.
/CR
