Masking Additional Attention Weights with Dropout

The purpose of dropout is to make the model robust. A pattern that survives when random parts of the network are removed is more likely to be a broad pattern than a dependence on one small group of units.

Programming

Interactive lab

Try it: Dropout at Training vs Inference

How dropout zeroes a random subset of a layer's outputs during training (with a visible seeded mask), how inverted dropout scales the survivors by 1 / (1 - rate) while inference is a no-op (or classic dropout scales at inference instead), and why the average training output matches the inference output.

How it works

For each unit draw u from the seeded generator; keep it when u >= rate, otherwise output 0.
Inverted dropout (Keras): multiply kept units by 1 / (1 - rate). Classic dropout: leave them unchanged.
At inference nothing is dropped: inverted dropout is a no-op, classic dropout multiplies by (1 - rate).
Average the training output over more and more masks: it converges to the inference output.

Default run (18 steps): A layer outputs 6 activations [0.8, 1.5, 0.2, 2, 1.1, 0.6]. Dropout rate 0.5, training mode, inverted dropout (as Keras does). … Average over 1000 masks (seeds 1..1000): [0.7936, 1.56, 0.198, 1.96, 1.1242, 0.5928]; largest gap to the inference output [0.8, 1.5, 0.2, 2, 1.1, 0.6] is 0.06. The average training output converges to the inference output: dropout preserves the expected activation.

Simplified: One activation vector with up to 10 units and a seeded teaching PRNG (mulberry32) instead of TensorFlow's generator.

Educational simulation

Loading the simulation…

A Model That Leans on One Narrow Path

Imagine a training process in which a model keeps finding one small group of internal units that carries most of the useful signal. Those units become a crutch. The model performs well on the examples it has already seen, but it has learned to depend on a narrow set of pathways rather than on a broad, robust pattern. This is the situation that the technique called dropout is designed to discourage.

In this article we first look at dropout as it appears in ordinary deep learning, and then move into the attention mechanism of transformer architectures such as GPT. The goal is to understand not only what dropout does, but exactly where it sits inside attention and why one placement is chosen over another.

Random Ignoring as a Training Device

Definition

Dropout is a training technique in which randomly selected hidden layer units are ignored. It helps prevent overfitting by stopping the model from relying too heavily on any specific set of units.

The word selection matters. The units are not chosen by a human, and they are not always the same units. Each training step draws a fresh random selection, so over many steps every unit is sometimes present and sometimes absent. No single unit can be guaranteed to be available, and therefore no single unit can become indispensable to the model. Because the set of available units keeps changing, the weights are shaped so that they do not depend on fragile combinations of units.

Key point

Training Mode Versus Inference Mode

Dropout is a training-time device. Random ignoring is useful while the model is learning, because it shapes the weights. Once training is finished and the model is used to produce outputs, dropout is disabled. The network then works with all of its units available. The same layer therefore behaves differently in the two situations: it drops units while training and passes everything through when the model is in use.

What changes between training mode and inference mode, and why does the same layer behave differently in each?

Trace the diagram from top to bottom. The top half is the training life of the layer, where units are ignored on each pass. The bottom half is the use of the finished model, where nothing is ignored. Dropout is not a permanent part of the computation; it is a mechanism that exists to shape the weights during learning.

Where Dropout Fits in Attention

Dropout is not limited to fully connected hidden layers. In the transformer architecture, including models like GPT, the attention mechanism can also use dropout. The source identifies two specific moments when this happens: after the attention weights have been calculated, or after those weights have been applied to the value vectors. In both cases, dropout is a random masking step inserted into the flow of attention.

The two placements answer a simple question: which quantity should be randomly ignored? Either the weights that decide how much each position attends to others, or the result obtained after those weights have been combined with the value vectors. The flow below shows the attention steps in order and marks both insertion points.

Where in the sequence of attention steps can dropout be inserted, and what does the data flow look like at each option?

Placement A sits between the weights and the value step, so the mask acts on the weights themselves. Placement B sits after the weights have been applied to the value vectors, so the mask acts on the combined result. The data passes through the same steps in both cases; only the point where the random masking happens is different.

Tracing a Dropout Mask on Attention Weights

Now we follow a dropout mask through placement A, where the mask is applied to the attention weights before they are applied to the value vectors. The example below uses invented numbers for four positions. It is a generated illustration of the masking step, not a figure taken from the source.

Which attention weights get zeroed by the mask, and what do the weights look like before and after dropout?

Worked example

Masking Four Attention Weights

Four attention weights for one query position are 0.40, 0.10, 0.30, and 0.20. A dropout mask marks positions 2 and 4 as ignored during this training step. Trace what happens to the weights before they are applied to the value vectors.

1
Start with the weights
The attention weights before masking are 0.40, 0.10, 0.30, and 0.20, one for each position.
2
Draw the mask
The random selection for this training step ignores positions 2 and 4. Positions 1 and 3 are kept. On another step a different selection could be drawn.
3
Apply the mask to the weights
Under placement A, the masked positions are set to zero, so position 2 becomes 0 and position 4 becomes 0. Positions 1 and 3 keep 0.40 and 0.30.
4
Pass the masked weights forward
The masked weights are then applied to the value vectors. Positions 2 and 4 contribute nothing to this step, so the output depends only on the kept positions.

Result

After masking, the weights are 0.40, 0, 0.30, and 0. The source pack describes the masking step but does not specify how the remaining values are scaled, so that detail is not covered here.

Notice what the mask does and does not do. It removes a random subset of the attention weights for this step, and the next training step may remove a different subset. Over many steps, every position is sometimes ignored, so no single attention pathway becomes indispensable. The mask is a training-time event: when the model is used for outputs, the masking is turned off.

Choosing Which Placement This Course Uses

The source names two valid times to apply dropout in attention and chooses one for its own walkthrough. Placing the two options side by side makes the choice easier to understand.

Question	Placement A: after the attention weights	Placement B: after the weights are applied to values
What is randomly ignored?	The weights that decide how much each position attends to others	The result after the weights are combined with the value vectors
Where in the flow is the mask?	Between the weight calculation and the value step	After the value step, before the output
Status in this course	Used in this course	Described by the source as a valid option, not the one used here
Why it is chosen	The source describes it as the more common variant in practice	Valid, but less common in practice according to the source

Comparing the two dropout placements in attention

Mistakes Learners Make With Dropout

Common mistake
Thinking dropout is always active
Why it fails
Dropout is used only during training and is disabled afterward, so it does not affect the model once it is in use.
Do this instead
Treat dropout as a training-time device. In inference mode, all units are available.
Common mistake
Believing the same units are always dropped
Why it fails
Each training step draws a fresh random selection, so the units ignored change from step to step.
Do this instead
Picture the mask as a new random draw on every step, not a fixed set of removed units.
Common mistake
Confusing the two placements in attention
Why it fails
Placement A masks the attention weights, while placement B masks the result after the weights are applied to value vectors.
Do this instead
Ask which quantity is being randomly ignored: the weights themselves, or the combined result.
Common mistake
Assuming the source gives a rescaling rule for the remaining weights; the source pack does not describe one, so do not treat a specific scaling method as taught here

Practice Tracing a Mask

EASY

Four attention weights are 0.25, 0.25, 0.25, and 0.25. A dropout mask under placement A ignores positions 1 and 3 on one training step. Write the weights after masking, then say which weights would be ignored on the next training step if the random selection changed.

Hints

Under placement A the mask acts on the attention weights, before they are applied to value vectors.
Masked positions contribute nothing to this step, so their weights are zero in the masked version.
Each training step draws a fresh random selection, so the next step can ignore a different set of positions.

Your attempt (not saved)

MEDIUM

Explain in two or three sentences why a model trained with dropout is less likely to depend on one narrow set of units, and why dropout is disabled once the model is used to produce outputs.