Lesson 75 found minimums by solving f′ = 0. That works for a parabola, but a neural network's loss has millions of inputs and no formula to solve. So training does something humbler and far more general: start somewhere, look at the gradient, take a small step downhill, and repeat until the loss stops falling. That's gradient descent, and it is how every model in this academy is trained, from a straight line to a large language model. This lesson takes its first steps by hand, writes it as a loop, shows what the learning rate does, and fits a line with it, checking the result against lesson 60's exact answer.
What you'll be able to do
Walk downhill to a low point: start with a guess, step against the gradient, repeat
Use the learning rate to set the step, and recognize steps that are too small, right, and too big
Write a training loop that updates variables inside a for loop and prints the loss
Fit a line by gradient descent, compare it with the exact answer from elimination, and train it again with PyTorch's zero_grad(), backward() and step()
Quick check
Have a guess before you start. You're on a foggy hillside and want to reach the valley floor. You can only feel the slope under your feet. What do you do?
Step downhill a little, feel again, and repeat
Walk in a straight line until you stop
Step uphill to get a better view
Stay still: you can't see
Show the answer
Step downhill a little, feel again, and repeat
Small steps in the downhill direction, checking the slope each time: that's gradient descent.
The math: step against the gradient, repeat
Purpose. Walk to a low point when no formula gives it.
Start with a guess θ (theta, standing for all the weights). Then repeat one rule:
θ←θ−η∇L(θ)
In plain words: new weights = old weights minus the learning rate times the gradient.
The gradient points uphill (lesson 78), so subtracting it moves downhill. The learning rate η scales the step. Near the bottom the gradient shrinks, so the steps shrink too, and the walk settles.
For L(x) = (x − 2)² + 1 from x = 0 with η = 0.25: the slope is −4, so x = 0 − 0.25 × (−4) = 1; then the slope is −2 and x = 1.5; then 1.75, then 1.875, closing in on the bottom at 2.
Quick check
From lesson 48: what is w − ηg for w = (1, 2), g = (4, −2) and η = 0.5?
(1, 3)
(−1, 3)
(3, 1)
(−3, 4)
Show the answer
(−1, 3)
ηg = (2, −1), and (1, 2) − (2, −1) = (−1, 3). Gradient descent is this update, repeated.
The picture. Step through four updates on the bowl (x − 2)² + 1, starting at x = 0 with η = 0.25. The picture draws the curve and the path down it: each step is an arrow from the old point to the new one, and the newest is lit. Watch the arrows shrink as the curve flattens.
Try it now
Then set the learning rate to 0.1 (slower), 0.5 (one step to the bottom), and 1.1 (the steps overshoot and the loss grows).
Symbols and formula. For one weight the rule is x ← x − η·L′(x); for many, each weight moves against its own partial, all at once:
w←w−η∂w∂L,b←b−η∂b∂L
In plain words: compute every partial first, then update every weight. Updating w before computing ∂L/∂b would compute that slope at a point the step never stood on.
The learning rate decides everything. Too small, and the walk crawls: with η = 0.1 the bowl's x goes 0.4, 0.72, 0.976, slowly. About right, and it settles quickly. Too big, past 2/L″ from lesson 76, and each step overshoots the bottom by more than it started: with η = 1.1, x goes 4.4, −0.88, 5.456, and the loss climbs.
Quick check
A training run's loss goes 5.0, 6.8, 9.3, 12.9. What is the likely problem?
The gradient is zero
The learning rate is too big
The learning rate is too small
The model is already perfect
Show the answer
The learning rate is too big
A loss that grows each step means every step jumps past the bottom by more than the last.
In your own words
In your own words: why does gradient descent slow down as it nears the bottom, even with a fixed learning rate?
Write a sentence or two, then compare it with a good answer in the lesson.
In Python: a training loop
Purpose. Repeat the update automatically, and watch the loss.
A training loop updates variables inside a for loop: each pass computes the slope, takes a step, and records the loss.
defloss(x):
return (x - 2) ** 2 + 1defslope(x):
return2 * (x - 2)
x = 0.0
lr = 0.25for step inrange(6):
x = x - lr * slope(x)
print(step + 1, round(x, 4), round(loss(x), 4))
It runs right here in your browser, with nothing to install.
The line x = x - lr * slope(x) reads the old x on the right and stores the new one on the left, so each pass starts where the last one ended. The loss falls on every step: 2, 1.25, 1.0625, …, toward the minimum value 1.
Code flow. Step through the same loop, three passes, one line at a time. Watch x, slope and loss: each pass reads the x the last one left, and the slope it finds is smaller each time.
A named mistake: updating one weight before computing the other's slope. With two weights, compute both partials from the same old values, then update both:
w, b = 0.0, 0.0
dw = 2 * (w - 1)
db = 2 * (b + 3)
w = w - 0.1 * dw
b = b - 0.1 * db
print(round(w, 4), round(b, 4))
If w were updated first and the slope for b then read the new w, the step would mix two different points. In this loss each slope only uses its own weight, so it happens not to matter; in a real model, where they're tangled together, it does.
Worked example
Worked example: finding a learning rate that diverges
Run 5 steps on the bowl at four learning rates, and print the final x and loss for each.
defloss(x):
return (x - 2) ** 2 + 1for lr in [0.05, 0.5, 0.9, 1.1]:
x = 0.0for step inrange(5):
x = x - lr * 2 * (x - 2)
print(lr, round(x, 4), round(loss(x), 4))
In the lesson you can change this and run it.
At 0.05 the walk crawls: after 5 steps x is only 0.819. At 0.5 it lands on the bottom at once. At 0.9 it overshoots back and forth, but each bounce is smaller, so it still settles near 2. At 1.1 every bounce is bigger than the last: x ends at 6.977 and the loss has grown to 25.77. The dividing line is η = 1, which is 2 ÷ L″ for this bowl.
Quick check
What does this print?
x = 4.0for step inrange(2):
x = x - 0.25 * 2 * (x - 2)
print(x)
5.0
2.5
2.0
3.0
Show the answer
2.5
From 4 the slope is 4, so x = 4 − 1 = 3; then the slope is 2, so x = 3 − 0.5 = 2.5.
Plot the loss. A list of losses is easier to read as a picture. This loop keeps the loss after each of 10 steps with η = 0.25, then draws it against the step number:
import matplotlib.pyplot as plt
x, losses = 0.0, []
for step inrange(10):
x = x - 0.25 * 2 * (x - 2)
losses.append((x - 2) ** 2 + 1)
plt.plot(range(1, 11), losses)
plt.xlabel("step")
plt.ylabel("loss")
plt.show()
Quick check
Read the plot. Where does the loss curve level off, and why there?
At 2, the first step's loss
At 1, the loss at the bottom, x = 2
At 0: every loss falls to 0
It never levels off: it keeps falling
Show the answer
At 1, the loss at the bottom, x = 2
(x − 2)² + 1 is smallest at x = 2, where it equals 1. The walk closes in on x = 2, so the loss closes in on 1, and the slope, and with it each step, shrinks toward 0.
Drill
A loop that descends
Fix the loop so it moves against the slope. Then set final to x after 10 steps from 5 with η = 0.2, rounded to 4 places, and diverges to the smallest rate in rates whose loss after 10 steps is bigger than the starting loss.
defloss(x):
return (x - 2) ** 2 + 1defrun(lr, steps):
x = 5.0for step inrange(steps):
x = x + lr * 2 * (x - 2)
return x
final = round(run(0.2, 10), 4)
rates = [0.3, 0.8, 1.05, 1.2]
diverges = 0print(final, diverges)
In the lesson, your code is checked as you write it.
In your own words
Why does η = 0.9 still settle, while η = 1.1 blows up, though both overshoot the bottom?
Write a sentence or two, then compare it with a good answer in the lesson.
In AI and ML: training a model
Purpose. This is how every model in the academy is trained.
Fit a line ŷ = w·x + b to four days of tent data from w = 0, b = 0. Each pass computes the errors, the gradient from lesson 78, and one update:
e=wx+b−y,w←w−ηn2e⋅x,b←b−ηn2∑e
In plain words: predict, measure the errors, compute both slopes, step both weights. Repeat.
import numpy as np
x = np.array([1.0, 2.0, 3.0, 4.0])
y = np.array([3.0, 5.0, 6.0, 9.0])
w, b = 0.0, 0.0for step inrange(1000):
e = w * x + b - y
dw = 2 / 4 * (e @ x)
db = 2 / 4 * np.sum(e)
w = w - 0.05 * dw
b = b - 0.05 * db
print(round(w, 4), round(b, 4), round(np.mean((w * x + b - y) ** 2), 4))
After 1,000 steps: w = 1.9, b = 1.0, loss 0.175. Lesson 60's exact method solves for the same weights in one go; gradient descent gets there by walking, which also works when no exact method exists.
Quick check
Why train networks by gradient descent instead of solving for the best weights directly, as lesson 60 did for a line?
Networks have no loss
Solving is only for integers
No equations to solve; descent needs only a gradient
Descent always gives the exact answer in a few steps
Show the answer
No equations to solve; descent needs only a gradient
The line's loss leads to linear equations; a network's doesn't. Walking downhill only needs the gradient, which any differentiable loss has.
In PyTorch. The same training run in the real tool. w and b are tensors that record how they're used (requires_grad=True, from lesson 72), and torch.optim.SGD is the optimizer that takes the steps:
import torch
x = torch.tensor([1.0, 2.0, 3.0, 4.0])
y = torch.tensor([3.0, 5.0, 6.0, 9.0])
w = torch.tensor(0.0, requires_grad=True)
b = torch.tensor(0.0, requires_grad=True)
opt = torch.optim.SGD([w, b], lr=0.05)
for step inrange(1000):
opt.zero_grad()
loss = ((w * x + b - y) ** 2).mean()
loss.backward()
opt.step()
print(round(w.item(), 4), round(b.item(), 4), round(((w * x + b - y) ** 2).mean().item(), 4))
It prints 1.9, 1.0 and 0.175, the NumPy cell's numbers. Read the two loops line for line:
NumPy loop
PyTorch loop
(nothing to clear)
opt.zero_grad(): clear the last step's slopes
e = w * x + b - y
loss = ((w * x + b - y) ** 2).mean()
dw = … and db = … by hand
loss.backward(): fills w.grad and b.grad
w = w - 0.05 * dw, and the same for b
opt.step(): every weight minus lr times its .grad
The learning rate 0.05 and the 1,000 steps are hyperparameters: settings you choose before training starts, unlike w and b, which training learns.
Quick check
A training loop leaves out opt.zero_grad(). What goes wrong?
The learning rate falls to 0
Nothing: backward() overwrites .grad each time
backward() adds to the old slopes, so they pile up
The loss is never computed
Show the answer
backward() adds to the old slopes, so they pile up
backward() adds its slopes into .grad. Without zero_grad(), step 2 uses the slopes of steps 1 and 2 added together, step 3 three of them, and so on, so the steps come out wrong.
What real training adds. Real data sets are too big to use every example on every step, so each step uses a small random batch: stochastic gradient descent, the S in SGD. Optimizers like momentum and Adam, in lesson 80, reuse past gradients to move faster through flat stretches and narrow valleys. Learning rates often start larger and shrink over time. But the heart of every one of them is the line you've written: weights minus learning rate times gradient.
Watching the loss is the habit. Printing the loss every so often is how you know training is working: it should fall quickly at first, then more slowly, then flatten (lesson 70). A loss that rises means the steps are too big; one that barely moves means they're too small, or the model has learned what it can. A plot of the loss against the step number, like the one you drew in the Python part, shows all three at a glance.
Quick check
Three training steps on the tent data from w = 0, b = 0, η = 0.05. What does this print?
import numpy as np
x = np.array([1.0, 2.0, 3.0, 4.0])
y = np.array([3.0, 5.0, 6.0, 9.0])
w, b = 0.0, 0.0
losses = []
for step inrange(3):
e = w * x + b - y
w = w - 0.05 * 2 / 4 * (e @ x)
b = b - 0.05 * 2 / 4 * np.sum(e)
losses.append(float(round(np.mean((w * x + b - y) ** 2), 4)))
print(losses)
[1.2134, 0.2186, 0.1911]
[37.75, 1.2134, 0.2186]
[1.2134, 1.2134, 1.2134]
[0.1911, 0.2186, 1.2134]
Show the answer
[1.2134, 0.2186, 0.1911]
The loss starts at 37.75 and falls fast: 1.2134, 0.2186, 0.1911. Most of the improvement comes in the first steps.
Drill
Train for 200 steps
Train the line on the data below for 200 steps from w = 0, b = 0, with η = 0.05. Set w_out, b_out and loss_out, each rounded to 3 places, and close to whether w is within 0.01 of the exact answer, 2.
import numpy as np
x = np.array([0.0, 1.0, 2.0, 3.0])
y = np.array([1.0, 3.0, 5.0, 7.0])
w, b = 0.0, 0.0for step inrange(200):
e = w * x + b - y
w = w - 0.05 * e.sum()
w_out = round(w, 3)
b_out = round(b, 3)
loss_out = 0
close = Falseprint(w_out, b_out, loss_out, close)
In the lesson, your code is checked as you write it.
In your own words
The exact method gives w = 1.9 at once, while gradient descent took 1,000 steps. Why is descent still the method that trains real models?
Write a sentence or two, then compare it with a good answer in the lesson.
31Formulas with fractions: clear them firstFreeClearing denominators, cross-multiplying, a letter in the bottom, and reciprocal formulas like 1/R = 1/R₁ + 1/R₂.35 min
37Inverse functions, lines and scalingFreeUndoing a function by making x the subject, domains that allow it, line forms, and unscaling predictions.40 min
38Sequences and sums: find n, a, d and rFreeEvery letter of the arithmetic and geometric formulas as the subject, and an average's window from its decay.40 min
40Review: rearrange anything (algebra)FreeBroken rearrangements to fix, mixed practice from the whole algebra block, a strategy chart, and its formula sheet.35 min
41Coordinates and distanceFreePoints as (x, y), Pythagoras, distance in any number of coordinates, and nearest neighbors.42 min
42Angles, degrees and radiansFreeMeasuring a turn, converting units, arcs, and the angle between two arrows.36 min
43Sine and cosineFreeAcross and up: right triangles, the unit circle, and directions and rotations.42 min
44Waves: height, speed and shiftFreeAmplitude, period and phase, sampling a wave, and position codes in transformers.37 min
68Vector equations: rearranging with arrowsFreeSolving for a vector or a scale, recovering a gradient from two weight snapshots, and an angle from a dot product.40 min
87Counting: choices, orders and groupsFreeThe counting principle, factorials, ordered picks and groups, and why attention's cost grows with the square of the length.43 min
88Probability: from counts to chancesFreeEqually likely outcomes, the complement, distributions, seeded simulation, and a next-word model.43 min
89Combining events: and, or, independentFreeThe or rule, independence, probability trees, drawing with and without replacement, and chains of steps.42 min
100Maximum likelihoodFreeLikelihoods as products, log-likelihoods as sums, grid search, underflow, and the negative log-likelihood loss.45 min
101Linear regression: the best lineFreeResiduals, least squares by formula, three roads to one line, the matrix form, and train versus test error.45 min
109One neuron, start to finishFreeWeighted sum, activation, loss, slopes and update, by hand and in code, and what one straight line can and can't split.45 min
111Training a small networkFreeThe backward pass in matrix form, gradient checks, a full training loop, and reading loss curves.45 min
112The same network in PyTorchFreenn.Sequential with the NumPy weights loaded, the same loss and slopes, then SGD and mini-batches.45 min
113Attention as matrix mathFreeScaled dot-product scores, the causal mask, a row-by-row softmax, and blending value vectors.45 min
114Residual connections and layer normFreeAdding the input back and why its slope keeps learning alive, layer norm by hand, and add and norm.40 min
115A transformer block, start to finishFreeA sentence through embeddings, attention, add and norm, a small network and next-word scores, every shape named.45 min
117Final test: transpose any formulaFreeSlips to fix from every topic, mixed practice of every kind of rearranging, the strategy chart and the full sheet.45 min
Learn the math behind AI from arithmetic up: numbers, algebra, vectors and matrices, calculus, probability and statistics, each shown as numbers moving and checked in Python, ending with a network you build and train from the formulas.
See the course →
Never trained an AI? Teach a small one to write, right in your browser. It takes about 20 seconds and needs no account. Teach it to write