Three lessons from three different points in the program, in the order a learner meets them.
Level 1 · The Disciple · 5 min
How a computer sees a picture
A photograph is a grid of numbers. Each square is a pixel, and in a greyscale image its value is how bright that square is, from 0 for black to 255 for white. A 28×28 handwritten digit is 784 of those values, and nothing else.
We start here on purpose. Every idea that follows, from a single prediction to a trained network, is arithmetic on a table like this one. Once you have seen the table, the mystery goes out of it.
What you will do: convert a shade of grey into a pixel value between 0 and 255, then count the numbers inside a 28×28 image and say what its shape tells you.
Level 1 · The Predictor · 6 min
Your first prediction in Python
Nothing hidden behind a library you have not met yet. You write the multiply-and-add yourself, first by hand and then in NumPy, and watch the number change as you change the weights.
import numpy as np
inputs = np.array([0.2, 0.8, 0.5]) # three features
weights = np.array([0.6, -0.3, 0.9]) # what the model has learned
bias = 0.1
total = inputs @ weights + bias # multiply, add, add the bias
prediction = 1 / (1 + np.exp(-total)) # squash the total into 0–1
print(round(prediction, 3)) # 0.745
Run it. Change one weight by 0.1 and run it again. Seeing the prediction move is the whole point of the lesson.
What you will do: change the weights until the prediction crosses 0.5, then explain which feature did the most work.
Level 3 · The Controller · 7 min
Why weights decide what matters
A weight is a dial. Turn it up and that input shouts; turn it down and it whispers; make it negative and it argues the other way. Most courses say “the model learns the weights” and move on. We spend a whole lesson on what that sentence means: how one large weight can swamp two small ones, why a bias lets a neuron fire even when its inputs are zero, and how to read a trained model and say out loud which inputs it actually listens to.
What you will do: load a trained model, rank its inputs by the size of their weights, and check your ranking against the data.
Read in ten minutes. The same three lessons inside the program come with a quiz, a lab and a marked mastery check. Back to the top ↑
