Attention (lesson 113) and a small network (lesson 110) are two pieces of a transformer block. Two more pieces glue them together, and both are short. The first adds each piece's input back onto its output: x + f(x). It looks like nothing, but its slope has a 1 in it, and that 1 is why networks dozens of layers deep can learn at all. The second rescales each word's vector to mean 0 and deviation 1, the standardizing of lesson 97 done one vector at a time, so the numbers stay a steady size from layer to layer. Together they're called "add and norm".
What you'll be able to do
Compute x + f(x) and its slope, 1 + f′(x), and say why the added 1 keeps slopes alive through many layers
Standardize one vector to mean 0 and deviation 1, then apply a learned scale γ and shift β
Tell layer norm from batch norm by which direction is averaged, and catch the slip of averaging over the batch
Follow the shapes through add and norm, and get the same numbers from PyTorch's nn.LayerNorm
Quick check
Have a guess before you start. Start with 1 and halve it, twenty times in a row. About how much is left?
About 0.000001
About 0.1
Exactly 0
10
Show the answer
About 0.000001
0.5²⁰ is about 0.00000095: halve anything twenty times and almost nothing is left. That's what happens to a slope passing back through twenty layers that each halve it, and this lesson's first idea is the fix.
The math: add the input back, then rescale
Purpose. Two short steps that let a network stack many layers: add each layer's input back, and keep every vector the same size.
Adding the input back. Any layer, a small network or attention, takes a vector x and gives back f(x). A residual connection sends x around the layer and adds it back on:
out=x+f(x)dxd(x+f(x))=1+f′(x)
In plain words: keep what came in, and add the layer's change to it. The slope is the slope of x, which is 1, plus the slope of f: the sum rule of lesson 73.
The layer now only has to learn a change, and the 1 in the slope turns out to matter most.
Quick check
From lesson 74: a path runs through three steps with slopes 2, 0.5 and 3. What is the slope of the whole path?
1
6
5.5
3
Show the answer
3
Along a chain, slopes multiply: 2 × 0.5 × 3 = 3. That's why many layers in a row can shrink a slope to almost nothing: every layer multiplies it again.
Why the 1 matters. A slope passing back through twenty layers is the product of their twenty slopes. A plain chain needs every slope close to 1 to survive, and that's hard to keep: twenty slopes of 0.5 leave 0.5²⁰, about 0.00000095, so the first layers barely learn. In a residual chain each factor is 1 + f′ instead, so each layer's own part only needs a slope close to 0. That's easy: a layer whose weights start small changes its input only a little. Twenty factors of 1.02 make about 1.49, a usable size. The 1 is not magic, though: twenty factors of 1.5 make about 3,325. Residual blocks work because each one's own part stays small.
The picture. Here is one residual layer, drawn as a calculation. Its own part halves its input, f(x) = 0.5x, and x = 2 goes two ways: through f, and around it. Press Next to send the slope back.
Try it now
The thick route around the layer carries the slope back multiplied by 1, whatever f does.
Rescaling each vector: layer norm. After an add, a vector's numbers can be any size, and they drift from layer to layer. Layer norm standardizes each vector on its own, as lesson 97 standardized a column: subtract its mean μ, divide by its deviation σ. Then it scales each entry by a learned γ (gamma) and shifts it by a learned β (beta), so training can give each entry back whatever size and center works best.
The picture. Step through add and norm for one word's vector, x = (2, 1, 1, 0), after a sublayer handed back f(x) = (2, 0, −1, 1). The bars are the vector's four entries, drawn on one scale for the whole flow, so you can watch them move: the add, the mean, the centering, the shrink to deviation 1, and last γ and β, which this layer has learned as (2, 1, 1, 1) and (0, 0, 0.5, 0).
Symbols and formula. For a vector y with d entries:
In plain words: center the vector on 0, shrink or stretch it to deviation 1, then let each entry take its own learned scale and shift.
ε (epsilon) is a tiny number, 0.00001 in PyTorch, so a vector whose entries are all equal, with σ = 0, never divides by 0. The flow above leaves it out to keep the hand arithmetic exact. γ and β have one number per entry, so a layer norm for vectors of 4 learns 8 numbers. A new one starts at γ = 1 and β = 0: pure standardizing, until training moves them.
Quick check
A vector is (5, 5, 5, 9). What is its mean, and what does its first entry become after subtracting it?
5 and 0
6 and 1
6 and −1
24 and −19
Show the answer
6 and −1
(5 + 5 + 5 + 9) ÷ 4 = 24 ÷ 4 = 6, and 5 − 6 = −1. The four centered entries, −1, −1, −1 and 3, add to 0.
In your own words
In your own words: why does adding x back make a deep network easier to train?
Write a sentence or two, then compare it with a good answer in the lesson.
In Python: slopes through twenty layers, and layer norm in three lines
Purpose. Measure the slope through twenty layers both ways, then write layer norm and check it.
defslope_through(f_slope, layers, residual):
total = 1.0for layer inrange(layers):
if residual:
total = total * (1 + f_slope)
else:
total = total * f_slope
return total
print(f"{slope_through(0.5, 20, False):.1e}", f"{slope_through(0.5, 20, True):.0f}")
print(f"{slope_through(0.02, 20, False):.1e}", f"{slope_through(0.02, 20, True):.2f}")
Each layer multiplies the slope again, by f′ in a plain chain or by 1 + f′ with a residual. With f′ = 0.5, the plain chain falls to 9.5e-07 and the residual one blows up to 3325. With f′ = 0.02, a layer that changes little, the plain chain is gone, 1.0e-34, while the residual one stays a usable 1.49. Residuals turn "keep every slope near 1", which is hard, into "keep every change small", which small starting weights already do.
Now layer norm, three lines on the flow's x + f(x). np.allclose asks whether two numbers agree to about five places, as math.isclose did in lesson 7:
The mean is 1.5, σ rounds to 1.5 (the 1e-5 under the root is ε), and ŷ is (1.6667, −0.3333, −1.0, −0.3333), the flow's numbers. True and True: mean 0 and deviation 1, up to float rounding.
Code flow. A model norms a whole batch of vectors at once, one per row. Step through three rows. axis=1 averages across each row, and keepdims=True keeps one number per row as a column, so it lines up with the row it came from. The second row is the first times ten.
A named mistake: averaging over the batch.axis=0 averages each column down the batch instead of across each vector. That's a real method, batch norm, used in networks for pictures, but it isn't layer norm: each word's numbers would then depend on whatever else is in the batch.
Across the row, the first vector comes out as before. Down the columns, it comes out as something else, set by the other two rows: change them and it changes too. A transformer reads one word at a time when it writes, with no batch to average over, so it norms each vector on its own.
Worked example
Worked example: the whole of add and norm
Write layer norm with its learned γ and β as a function, and run the flow's add and norm through it.
It prints (3.3333, −0.3333, −0.5, −0.3333), the flow's answer, and the shape (4,): a vector of 4 went in and a vector of 4 comes out. γ doubled the first entry and β lifted the third by 0.5. The order matters: add first, then norm, so the residual's sum is what gets rescaled.
Quick check
What does this print?
import numpy as np
v = np.array([3.0, 3.0, 5.0, 5.0])
print(v.mean(), v.std(), ((v - v.mean()) / v.std()).tolist())
4.0 1.0 [3.0, 3.0, 5.0, 5.0]
4.0 1.0 [-1.0, -1.0, 1.0, 1.0]
4.0 2.0 [-0.5, -0.5, 0.5, 0.5]
16.0 1.0 [-1.0, -1.0, 1.0, 1.0]
Show the answer
4.0 1.0 [-1.0, -1.0, 1.0, 1.0]
The mean is 4, every entry is 1 away from it, so σ = 1, and standardizing gives −1, −1, 1, 1.
Drill
Layer norm for a batch
Fix layer_norm(Y) so it standardizes each row of Y on its own: mean 0 and deviation 1 across the row, with 1e-5 under the square root. The starter has three slips.
import numpy as np
deflayer_norm(Y):
mu = Y.mean(axis=0)
var = ((Y - mu) ** 2).mean(axis=0)
return (Y - mu) / var
Y = np.array([[4.0, 1.0, 0.0, 1.0], [2.0, 2.0, 3.0, 1.0]])
out = layer_norm(Y)
print(out.round(4).tolist())
In the lesson, your code is checked as you write it.
In your own words
Why does layer norm average across each vector rather than down the batch?
Write a sentence or two, then compare it with a good answer in the lesson.
In AI and ML: add and norm in a transformer block
Purpose. Follow the shapes through add and norm, and run both halves in PyTorch.
Shapes through add and norm. A transformer block does add and norm twice: once after attention, once after the small network. Take the five camp-log words of lesson 113, "the fire was warm tonight", each 4 numbers wide: X is 5 × 4. Attention gives back one vector per word, 5 × 4; the add needs the two shapes equal, and gives 5 × 4; and layer norm rescales each row, each word, on its own: 5 × 4 again. Add and norm never change the shape, which is what lets a block stack on a block.
import numpy as np
X = np.array([[1.0, 0.0, 0.0, 0.0], [0.0, 1.0, 1.0, 0.0], [0.0, 0.0, 1.0, 0.0],
[0.0, 1.0, 0.0, 1.0], [1.0, 0.0, 0.0, 1.0]])
S = X @ X.T / np.sqrt(4)
S = np.where(np.tril(np.ones((5, 5))) == 1, S, -np.inf)
E = np.exp(S - S.max(axis=1, keepdims=True))
H = X + (E / E.sum(axis=1, keepdims=True)) @ X
mu = H.mean(axis=1, keepdims=True)
out = (H - mu) / np.sqrt(((H - mu) ** 2).mean(axis=1, keepdims=True) + 1e-5)
print(X.shape, H.shape, out.shape)
print(H[3].round(3).tolist(), out[3].round(4).tolist())
print(np.allclose(out.mean(axis=1), 0), out.std(axis=1).round(4).tolist())
"warm" was (0, 1, 0, 1); attention added (0.157, 0.686, 0.416, 0.427), so H's row is (0.157, 1.686, 0.416, 1.427). Layer norm turns that into (−1.1796, 1.1796, −0.78, 0.78). Every row now has mean 0 and deviation 1.0.
Quick check
X is 5 × 4. Suppose a sublayer gave back 5 × 16. Can a residual add the two?
Yes: NumPy stretches the smaller one
No: x + f(x) needs both the same shape
No: a residual only works after attention
Yes: the add makes it 5 × 20
Show the answer
No: x + f(x) needs both the same shape
The residual adds entry by entry, so f(x) must come back the same shape as x. That's why a transformer's small network widens to 16 and then narrows back to 4 before its add.
In PyTorch.nn.LayerNorm(4) is layer norm for vectors of 4: its weight is γ, its bias is β, and it adds ε = 1e-5 under the root. A residual connection is just a plus sign: x + layer(x). Here the layer is nn.Linear(1, 1, bias=False) with its one weight set to 0.5, so it halves its input, like the picture's f.
import torch
import torch.nn as nn
norm = nn.LayerNorm(4)
y = torch.tensor([2.0, 1.0, 1.0, 0.0]) + torch.tensor([2.0, 0.0, -1.0, 1.0])
print(norm(y).double().round(decimals=4).tolist())
with torch.no_grad():
norm.weight.copy_(torch.tensor([2.0, 1.0, 1.0, 1.0]))
norm.bias.copy_(torch.tensor([0.0, 0.0, 0.5, 0.0]))
print(norm(y).double().round(decimals=4).tolist())
layer = nn.Linear(1, 1, bias=False)
with torch.no_grad():
layer.weight.fill_(0.5)
x = torch.tensor([2.0], requires_grad=True)
out = x + layer(x)
out.backward()
print(layer(x).item(), out.item(), x.grad.item())
A new nn.LayerNorm starts at γ = 1 and β = 0, so its first answer is the three NumPy lines' (1.6667, −0.3333, −1.0, −0.3333). With the flow's γ and β loaded, it gives the flow's (3.3333, −0.3333, −0.5, −0.3333). Then autograd (lesson 82) finds the residual's slope for us: f(x) = 1.0, x + f(x) = 3.0, and x.grad = 1.5, which is 1 + f′. .double() turns float32 into float64 so the numbers round and print as NumPy's do.
The gradient reaching the first of twenty layers. Now real layers: twenty nn.Linear(4, 4) layers whose weights start small, made with NumPy's random generator so every run gets the same ones. The same layers are stacked twice, plain and with residuals, and backward() sends a slope from the top to the input x:
import numpy as np
import torch
import torch.nn as nn
rng = np.random.default_rng(0)
layers = []
for k inrange(20):
layer = nn.Linear(4, 4, bias=False)
with torch.no_grad():
layer.weight.copy_(torch.tensor(rng.normal(0, 0.1, (4, 4))))
layers.append(layer)
for residual in [False, True]:
x = torch.ones(4, requires_grad=True)
h = x
for layer in layers:
if residual:
h = h + layer(h)
else:
h = layer(h)
h.sum().backward()
print(residual, f"{x.grad.norm().item():.1e}")
x.grad.norm() is the size of the slope that reached x (lesson 49). Through the plain stack, the False line, it's 2.5e-15: nothing to learn from. Through the same layers with residuals, the True line, it's 1.3e+00, which is 1.3. One plus sign per layer is the whole difference.
Quick check
What does this print?
import torch
x = torch.tensor(3.0, requires_grad=True)
out = x + 0.25 * x
out.backward()
print(out.item(), x.grad.item())
3.75 0.25
0.75 1.25
3.75 1.25
3.75 1.0
Show the answer
3.75 1.25
out = 3 + 0.75 = 3.75, and its slope is 1 + 0.25 = 1.25: the 1 from x, the 0.25 from f(x) = 0.25x.
Drill
Add and norm for a sentence
Write add_norm(X, F) that adds the sublayer's output F to X and then layer-norms each row (1e-5 under the root). Then set shapes to the shapes of X, F and the result, and centered to whether every row of the result has mean 0 (use np.allclose).
import numpy as np
defadd_norm(X, F):
Y = (X - X.mean(axis=1, keepdims=True)) / X.std(axis=1, keepdims=True)
return Y + F
X = np.array([[1.0, 0.0, 0.0, 0.0], [0.0, 1.0, 1.0, 0.0], [0.0, 0.0, 1.0, 0.0]])
F = np.array([[0.5, 0.5, 0.0, 0.0], [0.2, 0.1, 0.4, 0.3], [0.0, 0.3, 0.3, 0.0]])
out = add_norm(X, F)
shapes = []
centered = Falseprint(out.round(4).tolist(), shapes, centered)
In the lesson, your code is checked as you write it.
In your own words
A transformer block is attention, add and norm, a small network, add and norm. Why can it be stacked dozens of times?
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