MISSION RECEIVEDAM 05:47
The moment people see a photo, they know whether it’s a cat or a dog. But what does a computer look at, and how does it decide? The order from HQ was short: Tell four pairs apart, and uncover the principle behind it. Starting with two fish and ending with chihuahuas and muffins, which confuse even people, you’ll solve four missions one by one together with the lab crew.
This is the log from three minutes after the alarm went off.
Click “Start Mission” to go to that episode. Mark a mission “Complete” when you finish it, and this device will remember.
Variables and Features: What’s the Difference?
Order Forty-nine fish brought in from the market have gotten mixed up. Sort the bream from the smelt using just two numbers: length and weight.
Gear Finding a one-line rule · perceptron · SVM · k-NN and standardization
A Photo Is a Matrix of Numbers
Order A single photo has arrived as thousands of numbers from 0 to 255. Find clues in that pile of numbers to tell a cat from a human.
Gear Pixel magnifier · the pixel-distance trap · convolution filters · facial landmarks
Stack Layers, Change the Coordinates
Order Both have fur, eyes, and ears. These two animals can’t be split by a single plane, so stack layers to tell them apart.
Gear 3D classification boundary · CNN architecture · softmax · Attentive CutMix
Low-Level Features Get Fooled, High-Level Features Decide
Order A round brown blob with three dark dots. Even people mix up this pair. Find out why the model gets fooled.
Gear Dark-spot filter · resolution experiment · pixel rules vs patterns · the model’s limits
From a pixel magnifier and a filter lab to a classification boundary you can rotate in 3D and CutMix that pastes patches together, you’ll see for yourself how the numbers actually change.
Each mission has code you can copy, download, or open directly as a notebook. It uses only Colab’s preinstalled libraries and built-in data, and the actual output is included.
Every section has a beginner-friendly explanation, tooltips on terms with a dotted underline, and check questions with explanations for wrong answers.
Prompts that use the code as a starting point for planning a hackathon, drafting a competition proposal, or designing experiments for a paper.
Analogies such as “a pattern stamp” and “electing a class representative,” along with notes on where each analogy breaks down.
Plans for lecture, on-the-job training (OJT), and 1:1 tutoring formats, questions for each section, and five common misconceptions.
◆ Go deeper into the fish mission — You can follow machine learning with the bream and smelt data (decision trees, k-NN, regression, cross-validation, and more) in detail in “Easy Machine Learning with Fish Market Data”.