This module adds no new math. Every formula in it was taught and practiced in Modules 1 to 7; what's new is running them together, the way a real system does. It starts with the smallest thing that learns: one neuron. Its training step is five familiar pieces, a weighted sum, an activation, a loss, slopes by the chain rule and an update, and once they run as one loop, the neuron learns. You'll take that step by hand, in NumPy and in PyTorch, then see exactly what one neuron can and cannot learn, which explains why networks stack them: the next lesson.
What you'll be able to do
Follow one neuron through a full training step by hand: weighted sum, activation, loss, slopes, update
Name the earlier lesson behind each of the five steps
Write the five steps as five lines of NumPy, then loop them until the loss reaches a target
Take the same step in PyTorch with backward() and optim.SGD, and get the hand numbers back
See what a single neuron can learn, a split of the plane by a straight line, and what it can't
Quick check
Have a guess before you start. A neuron gets inputs (2, 1), weights (0.5, −1) and bias 0.5. What is its weighted sum z = w·x + b?
0.5
2.5
1.5
−0.5
Show the answer
0.5
0.5 × 2 + (−1) × 1 + 0.5 = 1 − 1 + 0.5 = 0.5: a dot product plus the bias.
The math: five steps, all known
Purpose. Run one complete training step of one neuron by hand.
A neuron with two inputs has two weights w₁ and w₂ and a bias b. Here the inputs are x = (2, 1), the weights (0.5, −1), the bias 0.5, and the target, the answer it should give, is y = 1. One training step:
Weighted sum (lesson 50): z = w₁x₁ + w₂x₂ + b = 1 − 1 + 0.5 = 0.5.
Loss (lesson 79): L = (a − y)² = (0.5 − 1)² = 0.25.
Slopes (lessons 74 and 81): backward through each step, multiplying local slopes.
Update (lesson 79): every weight moves against its slope, by the learning rate.
Steps 1 to 3 are the forward pass, step 4 the backward pass. Step 4 is the only one with any work in it, and it's lesson 81's graph again: seven small steps, walked backward.
Quick check
From lesson 74: y depends on u, and u on x, with dy/du = −1 and du/dx = 2. What is dy/dx?
−2
1
2
−0.5
Show the answer
−2
The chain rule multiplies the links: −1 × 2 = −2.
The picture. The flow draws the neuron as that graph. Six circles start with a number: the inputs and weights on the left, the bias b, which joins at z, and the target y, which joins at e. Each empty circle is one small step: p₁ and p₂ multiply an input by its weight, s adds them, z adds the bias, a is the relu, e is the error a − y, and L squares it. Forward, each value is written in its circle. Backward, the slope of L for each value is written under its circle after a ∂, and the arrow that slope came back along keeps the local slope it was multiplied by, after a ×. On the PyTorch tab, the whole backward pass is one call, L.backward().
Try it now
With b = −1.5 instead, z would be negative, relu would give 0, and every weight's slope would be 0: a "dead" neuron learns nothing from this example.
Symbols and formula. For a relu neuron with squared loss, the backward pass collapses to:
∂z∂L=2(a−y)⋅[z>0]∂wj∂L=∂z∂Lxj∂b∂L=∂z∂L
In plain words: the loss pushes back by twice the error, relu passes the push only if the neuron was on, and each weight's share is the push times its own input.
The update. With learning rate 0.1, each weight and the bias step against their slopes: w₁ = 0.5 − 0.1 × (−2) = 0.7, w₂ = −1 − 0.1 × (−1) = −0.9, and b = 0.5 − 0.1 × (−1) = 0.6. Here is the whole step on the neuron drawn as one unit of a network, the way the next lessons draw it. Its third input is always 1, so the weight on that arrow is the bias: that's why ∂L/∂b is ∂L/∂z times 1. Press Next for the backward pass, then the update.
Quick check
∂L/∂z = −1 for an example with inputs (3, 0). What are the slopes of w₁ and w₂?
3 and 0
−1 and −1
−3 and −1
−3 and 0
Show the answer
−3 and 0
Each weight's slope is ∂L/∂z times its input: −1 × 3 and −1 × 0. A zero input means its weight isn't touched.
In your own words
In your own words: why does a neuron with a negative z, under relu, learn nothing from that example?
Write a sentence or two, then compare it with a good answer in the lesson.
In Python: five lines, then a loop
Purpose. Write the five steps as code, then repeat them until the loss is small.
import numpy as np
x = np.array([2.0, 1.0])
w = np.array([0.5, -1.0])
b = 0.5
y = 1.0
z = float(w @ x + b)
a = max(z, 0.0)
L = (a - y) ** 2
dz = 2 * (a - y) * (z > 0)
w = w - 0.1 * dz * x
b = b - 0.1 * dz
print(L, w.tolist(), round(b, 4))
print(round((max(float(w @ x + b), 0.0) - y) ** 2, 4))
From z on, the lines are the five steps: sum, activation, loss, slope, and the update, which takes two lines, one for w and one for b. float() makes z a plain Python number instead of a NumPy one. (z > 0) is True or False, which counts as 1 or 0: relu's slope. The loss falls from 0.25 to 0.01 after one step, matching the hand calculation.
Code flow. Step through the same lines one at a time and watch the variables. z, a and L appear in the forward pass; dz is the push the backward pass sends back to z; and the last two lines move w and b. w shows as two boxes, the two weights from the math part's pictures.
A named mistake: updating one weight before computing the others' slopes. If w₁ is updated first and the slope of w₂ is then computed from the new w₁, the two slopes come from different points on the loss surface. Compute every slope from the same forward pass, then update all the weights together, as lesson 79's loop did.
On real data the same lines run over a batch: X @ w + b gives every example's z at once, and the slopes are averaged over the batch.
Worked example
Worked example: train until the loss reaches a target
Four examples whose targets follow y = 0.5x₁ + 0.5x₂. Train one relu neuron on all four at every step until the mean squared loss is below 0.001.
import numpy as np
X = np.array([[2.0, 1.0], [1.0, 3.0], [3.0, 2.0], [0.0, 1.0]])
y = np.array([1.5, 2.0, 2.5, 0.5])
w = np.array([0.1, 0.1])
b = 0.0
first = -1for step inrange(200):
z = X @ w + b
a = np.maximum(z, 0)
L = np.mean((a - y) ** 2)
if step < 3:
print(step, round(float(L), 3))
if L < 0.001and first == -1:
first = step
dz = 2 * (a - y) * (z > 0) / 4
w = w - 0.05 * (X.T @ dz)
b = b - 0.05 * np.sum(dz)
print(first, w.round(2).tolist(), round(float(b), 2))
In the lesson you can change this and run it.
The loss falls fast at first, from 2.04 to 0.162 to 0.018 in two steps, then slowly, and first drops below 0.001 at step 60. The weights head toward 0.5 and 0.5, the rule the targets were made with. Two details are worth seeing: X.T @ dz sums each weight's slope over the four examples in one product (lesson 56), and the division by 4 averages them, so the learning rate doesn't have to change with the batch size.
Quick check
What does this print?
import numpy as np
x = np.array([3.0, 0.0])
w = np.array([1.0, 1.0])
z = w @ x - 1
dz = 2 * (max(z, 0.0) - 1) * (z > 0)
print(z, dz, (dz * x).tolist())
3.0 4.0 [12.0, 0.0]
2.0 0.0 [0.0, 0.0]
2.0 2.0 [6.0, 0.0]
2.0 2.0 [6.0, 2.0]
Show the answer
2.0 2.0 [6.0, 0.0]
z = 3 − 1 = 2, relu keeps it, so ∂L/∂z = 2(2 − 1) = 2, and the weight slopes are 2 × 3 and 2 × 0.
Drill
Train to a target loss
Fill in the slope dz and the two update lines (learning rate 0.05, slopes averaged over the 4 examples), then set first to the first step whose loss is below 0.001, and w_final to the weights, rounded to 1 place, as a list.
import numpy as np
X = np.array([[2.0, 1.0], [1.0, 3.0], [3.0, 2.0], [0.0, 1.0]])
y = np.array([1.5, 2.0, 2.5, 0.5])
w = np.array([0.1, 0.1])
b = 0.0
first = -1for step inrange(200):
z = X @ w + b
a = np.maximum(z, 0)
L = np.mean((a - y) ** 2)
if L < 0.001and first == -1:
first = step
dz = 0
w_final = w.round(1).tolist()
print(first, w_final)
In the lesson, your code is checked as you write it.
In your own words
The loss fell from 2.04 to 0.018 in two steps, then took 58 more to get below 0.001. Why does progress slow down?
Write a sentence or two, then compare it with a good answer in the lesson.
In AI and ML: the same step in PyTorch, and one neuron's limit
Purpose. Take the step with PyTorch's own tools, then find the limit of a single neuron, which is why networks have layers.
In PyTorch. Each node in that row has a PyTorch name, and one training step uses them in order. The weights are tensors made with requires_grad=True, so PyTorch records the forward pass as it runs (lesson 82). torch.optim.SGD is given the list of weights it may move and the learning rate, as in lesson 80. Then zero_grad() clears any old slopes, L.backward() puts each weight's slope in its .grad, and step() moves every weight against its slope:
import torch
x = torch.tensor([2.0, 1.0])
y = 1.0
w = torch.tensor([0.5, -1.0], requires_grad=True)
b = torch.tensor(0.5, requires_grad=True)
opt = torch.optim.SGD([w, b], lr=0.1)
z = w @ x + b
a = torch.relu(z)
L = (a - y) ** 2
opt.zero_grad()
L.backward()
print(L.item(), w.grad, b.grad)
opt.step()
print(w, b)
with torch.no_grad():
print(round(((torch.relu(w @ x + b) - y) ** 2).item(), 4))
Every number is one you found by hand: the loss 0.25 and the slopes −2, −1 and −1 from the flow, then the new weights 0.7, −0.9 and 0.6 and the new loss 0.01, which the NumPy cell printed too. with torch.no_grad(): keeps that last check from being recorded, since it isn't part of training. A real model hands the optimizer all its weights at once, optim.SGD(model.parameters(), lr=0.1), and repeats the three calls in a loop: lesson 111.
Quick check
What does this print?
import torch
w = torch.tensor([1.0, 2.0], requires_grad=True)
x = torch.tensor([3.0, -1.0])
z = w @ x
z.backward()
print(z.item(), w.grad)
5.0 tensor([ 3., -1.])
1.0 tensor([3., 1.])
1.0 tensor([ 3., -1.])
1.0 tensor([1., 2.])
Show the answer
1.0 tensor([ 3., -1.])
z = 1 × 3 + 2 × (−1) = 1, and z's slope for each weight is that weight's own input: 3 and −1. It's ∂z/∂wⱼ = xⱼ, the rule behind every weight's slope in this lesson.
What one neuron can learn. A sigmoid neuron with two inputs is lesson 102's logistic regression: its decision boundary is where w₁x₁ + w₂x₂ + b = 0, a straight line across the plane of inputs. Whatever it learns, it can only say "yes on this side, no on that side".
Take the trail club's signs, sorted by two yes-or-no features: pointed and wide. The four kinds sit at the four corners of the plane: a plain blaze (0, 0), a tent (0, 1), an arrow (1, 0) and a summit sign (1, 1). The neuron below has its weights set by hand to spot summit signs.
AND fits. "Summit sign?" is AND: yes only when both features are 1. Training finds weights for it by itself. Below is the worked example's loop with a sigmoid in place of relu; with the log loss, the slope of the loss for z is simply p − y (lesson 102). The rows of X are the blaze, the tent, the arrow and the summit sign.
import numpy as np
X = np.array([[0.0, 0.0], [0.0, 1.0], [1.0, 0.0], [1.0, 1.0]])
y = np.array([0.0, 0.0, 0.0, 1.0])
w = np.zeros(2)
b = 0.0for step inrange(5000):
p = 1 / (1 + np.exp(-(X @ w + b)))
w = w - 1.0 * (X.T @ (p - y)) / 4
b = b - 1.0 * np.mean(p - y)
p = 1 / (1 + np.exp(-(X @ w + b)))
loss = np.mean(-(y * np.log(p) + (1 - y) * np.log(1 - p)))
print(w.round(1).tolist(), round(b, 1), round(float(loss), 3))
print(p.round(3).tolist())
Training finds weights near (10.7, 10.7) and bias −16.2: only the summit sign scores above 0, so only it gets a high chance, 0.994. All four come out right, with a loss of 0.003. It's the same kind of line as the hand-set neuron's; the bigger weights just make its answers surer.
Quick check
What does this print? The AND weights from above, on two new signs: a pointed one only half as wide as a summit sign, and a summit sign.
import numpy as np
X = np.array([[1.0, 0.5], [1.0, 1.0]])
w = np.array([10.7, 10.7])
b = -16.2print((1 / (1 + np.exp(-(X @ w + b)))).round(2).tolist())
[0.0, 0.99]
[0.5, 0.99]
[0.99, 0.46]
[0.46, 0.99]
Show the answer
[0.46, 0.99]
(1, 0.5) scores 10.7 + 5.35 − 16.2 = −0.15, just below 0, so its chance is just under 0.5; (1, 1) scores 5.2 and gets 0.99. Close to the line, the neuron is unsure.
XOR doesn't. Now ask "arrow or tent?", one feature but not both: yes for (1, 0) and (0, 1), no for (0, 0) and (1, 1). That's XOR, and no straight line puts the two yeses on one side and the two noes on the other. Training does the best it can, and the best is no line at all: the weights stay at 0, every prediction is 0.5, the loss is stuck at 0.693 (log 2), and accuracy at threshold 0.5 is 50%, a coin toss. This is not a training failure; it's a limit of what one line can draw.
Quick check
Why can no single neuron learn XOR?
One neuron draws one line, and no line splits XOR
XOR needs negative inputs
The learning rate is always too small for XOR
XOR has too few examples
Show the answer
One neuron draws one line, and no line splits XOR
The yeses sit at opposite corners, (0, 1) and (1, 0); any line with both on one side also has a no there.
Two lines make XOR. Two neurons in a hidden layer can each draw their own line. In the network below, OR says yes when at least one feature is 1 and NAND says yes unless both are; the output neuron says yes only when both of them do. Here a is pointed and b is wide. The weights are set by hand, and lesson 110 builds networks like it.
Drill
A data set it fits, and one it can't
Train the sigmoid neuron on AND, the summit-sign question, and on XOR, arrow or tent (5,000 steps, learning rate 1). Set acc_and and acc_xor to the accuracy at threshold 0.5, and loss_xor to XOR's final log loss, rounded to 3.
import numpy as np
defsigmoid(z):
return1 / (1 + np.exp(-z))
deftrain(X, y):
w = np.zeros(2)
b = 0.0for step inrange(5000):
p = sigmoid(X @ w + b)
w = w - 1.0 * (X.T @ (p - y)) / 4
b = b - 1.0 * np.mean(p - y)
return w, b
X = np.array([[0.0, 0.0], [0.0, 1.0], [1.0, 0.0], [1.0, 1.0]])
y_and = np.array([0.0, 0.0, 0.0, 1.0])
y_xor = np.array([0.0, 1.0, 1.0, 0.0])
acc_and = 0
acc_xor = 0
loss_xor = 0print(acc_and, acc_xor, loss_xor)
In the lesson, your code is checked as you write it.
In your own words
XOR's loss stopped at 0.693 with every prediction 0.5. Why is that the best a single line can do with log loss?
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