How machines learn from mistakes
Take this if you have heard that AI models are trained and want to know what the word actually describes.
What you’ll be able to do
You can explain what a model does after it gets an answer wrong, and predict what happens when the learning rate is too big or too small.
What it assumes
You can read a graph with an x and y axis. No calculus, no coding, nothing installed.
- Price
- Free
- Length
- 1 week
- Sessions
- 4
- Effort
- ~2 h/week · 2 h total
- For
- Anyone
Adds it to My courses on your dashboard. No card, nothing to cancel.
What’s taught, in order
- 01
Error as a landscape
Before any algorithm, the idea that a wrong answer has a size, and that all the possible wrong answers form a surface with a lowest point. This is the mental picture everything else hangs off.
- 02
Which way is downhill
Working out the direction of steepest descent by hand at three points on a curve, so that when the word gradient arrives it is already a thing you have measured rather than a symbol.
- 03
Step size, and why it ruins everything
Driving the visualiser at a learning rate of 0.01, then 0.15, then 1.9, and watching the same algorithm crawl, converge and then fly apart. The failure modes are the lesson.
- 04
Where this shows up again
The same loop, unchanged, is what trains a language model with billions of parameters. Seeing that the algorithm does not get more complicated is the point of ending here.
What you build
Using the free gradient descent visualiser, get from the starting point to the minimum in fewer steps than the automatic solver takes. Write down the learning rate you settled on, one rate that was too small and one that diverged, and what the path looked like in each case.
Check it yourself against this
- You beat the automatic solver's step count at least once.
- You can name a learning rate that converged and one that diverged.
- You can describe, in one sentence, what the algorithm does when it overshoots.
- You noticed that a smaller step is not always a slower result, and can say why.
Then answer this
Why does an algorithm that only ever looks at the slope directly beneath it still find the bottom of the bowl?
Two or three sentences, in your own words. If you cannot, the course is not finished — go back to the module it came from.
What this course does not do
This course does not teach you to write the algorithm in code, and it stays on smooth single-variable curves where descent always works. Real training surfaces have ravines, plateaus and local minima that break the intuition you build here, which is what the Foundations phase course exists to fix.
Where this leads
The interactive visualiser at the centre of this course is free and needs no signup, so you can go and feel the learning-rate trade-off for yourself before deciding whether to start the course at all.
FAQ
- Do I need to know calculus to take this?
- No. The course measures slopes by hand off a graph rather than differentiating anything, which is deliberate — the idea of a gradient is geometric before it is algebraic, and meeting it in that order makes the calculus easier later rather than harder.
- Is this really free, or free until a paywall?
- Genuinely free, all four sessions, and you do not need to enter a card. Creating an account only matters if you want the completion certificate, which is signed and publicly verifiable, or want your progress saved between sessions.
- How is this different from watching a video about gradient descent?
- You never watch anything. Every session hands you a live visualiser and asks you to predict what a change will do before you make it, then shows you whether you were right. The prediction step is the part that makes it stick.
Other courses
This is a standalone course. It does not add weeks to the 78-week curriculum or change your roadmap. If you want the whole path instead, the full programme is ₹399/month · ₹2,000/year — see what it covers.