See like a CNN
Take this if you want to know how an image becomes a prediction, filter by filter, rather than layer diagram by layer diagram.
What you’ll be able to do
You can hand-design a convolution kernel for a specific effect, explain what a feature map represents, and train a small convolutional network on image data you produced yourself.
What it assumes
Comfortable Python, and having met the idea of a neural network — the free neural network course is enough.
- Price
- ₹399
- Length
- 2 weeks
- Sessions
- 8
- Effort
- ~5 h/week · 10 h total
- For
- Class 11–12 · College · Working professional
Adds it to My courses on your dashboard so you can find it again. Payment is not open yet — nothing is charged.
What’s taught, in order
- 01
A kernel is a small question
Sliding a three-by-three grid of numbers over an image and computing by hand what comes out. Designing an edge detector yourself means convolution is arithmetic you have done, not an operation you invoke.
- 02
Feature maps and what fires
Visualising which parts of an image excite which filter, and finding that early layers respond to edges and later ones to shapes. This is the hierarchy claim, checked rather than asserted.
- 03
Pooling, stride and why size shrinks
Where the spatial resolution goes and what is traded for it. Being able to compute an output shape by hand saves hours of debugging once you are stacking real layers.
- 04
Training on your own images
Assembling a small dataset, training a compact network on it, and confronting the gap between a model that works on clean data and one that works on yours.
What you build
Photograph or scan your own handwritten digits, at least twenty examples of each of three digits, and train a small convolutional network to classify them. Report its accuracy, the digit pairs it confuses, and what changed when you added the examples written with a different pen or in worse light.
Check it yourself against this
- The dataset is your own handwriting, with at least sixty labelled images.
- You report accuracy on images the model was not trained on.
- You identify which digits get confused and offer a visual reason.
- You tested robustness by varying pen, paper or lighting, and reported the effect.
- You can hand-compute the output shape of your first convolutional layer.
Then answer this
Nobody told your network what an edge is, yet its first layer learned edge detectors anyway — why is that the natural thing for it to learn first?
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
Small images, a small network, and short training runs that fit in a browser session. Modern architectures, transfer learning, detection and segmentation are all beyond this sprint, and the accuracy you reach will be well below what a pretrained model would give you.
Where this leads
The convolution visualiser is free on its own page, and hand-editing a three-by-three kernel to see an edge detector appear is worth doing before you decide whether this sprint is for you.
FAQ
- Do I need a GPU?
- No. The networks are deliberately small and the images are low resolution, so everything trains in your browser in a couple of minutes. That constraint is also pedagogically useful — a model you can retrain in two minutes is a model you will actually experiment with.
- Is PyTorch used?
- Lightly, after you have implemented a convolution yourself with loops. Writing the slow version first is what makes the fast version legible rather than magical.
- Why handwriting rather than a standard dataset?
- Because a standard dataset is pre-cleaned and hides the real lesson. Your own handwriting under your own lighting produces exactly the messy failure cases that teach you what a vision model is actually sensitive to.
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.