Early stopping and checkpoints quiz

Question 1

What is early stopping in machine learning?

Accepted Answer

A regularization method that halts training when validation performance stops improving

Answer

A technique to stop training when training loss reaches zero

Answer

A method to reduce the learning rate during training

Answer

A technique to stop training after a fixed number of epochs regardless of performance

Question 2

Which metric is typically monitored during early stopping, and why?

Accepted Answer

Validation loss

Answer

Training accuracy

Answer

Training loss

Answer

Learning rate

Question 3

What does the 'patience' parameter mean in early stopping?

Accepted Answer

The number of consecutive non-improving epochs allowed before stopping training

Answer

The number of epochs to train before checking for improvement

Answer

The minimum improvement required to reset the counter

Answer

The total maximum number of training epochs permitted

Question 4

What is a model checkpoint in deep learning?

Accepted Answer

A saved snapshot of model weights at a specific point during training

Answer

A technique to reduce the learning rate at certain training steps

Answer

A validation score threshold that triggers training to stop

Answer

A type of regularization applied at the end of each epoch

Question 5

Why must early stopping use a validation set rather than the test set?

Accepted Answer

Using the test set for stopping decisions would bias the final evaluation

Answer

The test set is always too small for reliable loss computation

Answer

Validation loss is always lower than test loss by definition

Answer

Test sets do not support computing loss in most frameworks

Question 6

How does early stopping act as a form of regularization?

Accepted Answer

It limits weight updates, keeping weights closer to initialization and constraining model complexity

Answer

It directly penalizes large weights by adding a term to the loss function

Answer

It increases the learning rate to prevent the model from memorizing data

Answer

It randomly drops neurons during training to reduce co-adaptation

Question 7

What problem can arise if the patience parameter is set too low, such as patience=1?

Accepted Answer

Training will stop too early due to temporary fluctuations in validation loss

Answer

The model will overfit because it trains too many extra epochs

Answer

The checkpoint will not save the best weights to disk

Answer

The optimizer will diverge because the learning rate is not reduced

Question 8

What is the key difference between saving the 'last' model versus the 'best' model checkpoint?

Accepted Answer

The best checkpoint saves at peak validation performance, while the last saves at the final epoch which may be overfitted

Answer

The last checkpoint always has higher accuracy than the best checkpoint

Answer

The best checkpoint saves optimizer state while the last checkpoint saves only weights

Answer

Both are always identical when early stopping is used correctly

Question 9

In a training learning curve, what pattern signals that early stopping should have triggered earlier?

Accepted Answer

Training loss decreases while validation loss starts increasing (diverging curves)

Answer

Both training and validation loss are high and decreasing together at the same rate

Answer

Training loss is increasing while validation loss remains stable

Answer

Validation loss reaches zero before training loss does

Question 10

What does the Keras EarlyStopping parameter restore_best_weights=True do?

Accepted Answer

After stopping, it reverts the model's weights to the epoch with the best monitored metric

Answer

It retrains the model from scratch using the best hyperparameters found

Answer

It saves the best weights to disk at every epoch automatically

Answer

It restores the initial random weights and retrains with a lower learning rate

Question 11

Which of the following Keras EarlyStopping configurations is correct for monitoring validation loss with patience 10 and restoring best weights?

Accepted Answer

Option A — monitor='val_loss', patience=10, restore_best_weights=True

Answer

Option B — monitor='loss' (training loss), patience=10, restore_best_weights=False

Answer

Option C — monitor='val_accuracy', patience=0, restore_best_weights=True

Answer

Option D — monitor='val_loss', patience=1, restore_best_weights=True

Question 12

What does the min_delta parameter do in Keras EarlyStopping?

Accepted Answer

It defines the minimum change required in the monitored metric to count as an improvement

Answer

It sets the minimum number of epochs before early stopping can activate

Answer

It sets the minimum learning rate allowed during the training run

Answer

It specifies the minimum validation accuracy required before saving a checkpoint

Question 13

What does the following Keras ModelCheckpoint code do?

Accepted Answer

Saves the model only when validation loss reaches a new minimum, always keeping the best model

Answer

Saves the model at every epoch, overwriting the previous file each time

Answer

Saves the model only when validation accuracy reaches a new maximum

Answer

Saves the model at every 5th epoch with the current weights

Question 14

How is early stopping different from learning rate scheduling with ReduceLROnPlateau?

Accepted Answer

Early stopping terminates training when metrics plateau; scheduling reduces the learning rate to continue optimizing

Answer

Early stopping reduces the learning rate; scheduling terminates training

Answer

Both are identical — learning rate scheduling is just a renamed form of early stopping

Answer

Learning rate scheduling monitors validation loss while early stopping monitors training loss

Question 15

Why does training loss keep decreasing while validation loss starts to increase?

Accepted Answer

The model is overfitting: memorizing training-specific patterns that do not generalize to validation data

Answer

The model's learning rate is too low, causing slow convergence on training data only

Answer

The validation set is too large, making its loss artificially inflated

Answer

The optimizer is alternating updates between training and validation data randomly

Question 16

Under what mathematical condition is early stopping in gradient descent approximately equivalent to L2 regularization?

Accepted Answer

For quadratic loss with gradient descent, where training steps T acts as inverse regularization strength (λ ≈ 1/T)

Answer

When the learning rate equals exactly 1.0 and the loss function is cross-entropy

Answer

Only when batch size equals the full training set size and momentum is zero

Answer

Only when dropout is also applied simultaneously with early stopping

Question 17

How does early stopping address the bias-variance trade-off in machine learning?

Accepted Answer

It reduces variance (overfitting) by stopping before the model memorizes training data, at the cost of slightly higher bias

Answer

Early stopping eliminates both bias and variance by always training to full convergence

Answer

It eliminates bias by ensuring the model always reaches the global minimum of training loss

Answer

Early stopping increases variance intentionally to improve training speed

Question 18

In which scenario might early stopping actually hurt model performance?

Accepted Answer

When the validation curve is noisy and the true minimum occurs late, causing premature stopping

Answer

When training on a very large dataset with millions of samples and a simple model

Answer

When the model architecture is too simple to overfit the training data

Answer

When using batch normalization with very large batch sizes

Question 19

What does setting mode='max' versus mode='min' control in Keras EarlyStopping?

Accepted Answer

mode='min' means lower metric values are better (loss); mode='max' means higher values are better (accuracy)

Answer

mode='max' trains for more total epochs while mode='min' trains for fewer epochs

Answer

mode='max' saves the last checkpoint; mode='min' saves only the best checkpoint

Answer

mode refers to the optimizer: 'max' selects Adam, 'min' selects SGD

Question 20

At which point during training is the early stopping criterion evaluated?

Accepted Answer

At the end of each epoch after evaluating the validation set

Answer

After every mini-batch gradient update during training

Answer

At the very start of training before any weights are updated

Answer

Every 100 gradient update steps regardless of epoch boundaries

Question 21

Why is checkpoint frequency an important design decision in long training runs?

Accepted Answer

Too-infrequent checkpointing risks losing training progress on crash; too-frequent creates large storage overhead

Answer

Frequent checkpoints improve model accuracy by providing additional gradient signals

Answer

Checkpoint frequency only matters when using learning rate scheduling callbacks

Answer

All deep learning frameworks automatically checkpoint at the optimal frequency

Question 22

How can early stopping interact unexpectedly with batch normalization layers?

Accepted Answer

Early stopping can trigger before batch norm running statistics stabilize, causing degraded inference performance

Answer

Batch normalization prevents overfitting completely, making early stopping redundant

Answer

Batch normalization and early stopping cannot be used in the same model architecture

Answer

Early stopping speeds up batch normalization convergence by reducing total training steps

Question 23

What is the risk of using training loss instead of validation loss as the early stopping criterion?

Accepted Answer

Training loss almost always keeps decreasing, so early stopping would never trigger and the model could overfit

Answer

Training loss reaches zero faster than validation loss, causing premature stopping

Answer

The model would stop training too early because training loss decreases quickly at the start

Answer

Training loss is not available as a callback metric in most frameworks like Keras

Question 24

What should replace ??? in the PyTorch early stopping code to correctly save the best model when a new minimum validation loss is found?

Accepted Answer

torch.save(model.state_dict(), 'best_model.pth')

Answer

model.eval() followed by model.freeze()

Answer

optimizer.zero_grad()

Answer

torch.save(optimizer.state_dict(), 'best_model.pth')

Question 25

What is the combined effect of using both EarlyStopping and ReduceLROnPlateau callbacks in Keras?

Accepted Answer

ReduceLROnPlateau first tries smaller learning steps at a plateau; EarlyStopping stops if improvement still doesn't occur

Answer

They conflict with each other and should never be used together in the same training run

Answer

Both callbacks monitor different metrics and always trigger at exactly the same epoch

Answer

ReduceLROnPlateau replaces the need for EarlyStopping entirely when used correctly

Question 26

What happens to the model's weights immediately after EarlyStopping triggers?

Accepted Answer

Training halts; if restore_best_weights=True weights revert to the best epoch, otherwise the last epoch's weights are kept

Answer

The model resets to random initial weights and training restarts from scratch

Answer

Training continues but at a much-reduced learning rate automatically

Answer

The model is automatically exported to disk in all popular formats regardless of settings

Question 27

How do learning curves from exploratory training runs guide your choice of patience value?

Accepted Answer

Learning curves reveal how quickly validation loss recovers from dips, guiding patience selection

Answer

Learning curves show the optimal batch size, which indirectly determines the correct patience value

Answer

Patience should always be set to the exact epoch count where training loss first plateaus

Answer

Learning curves have no relationship to patience; it must be determined by grid search only

Question 28

How does early stopping compare to dropout as a regularization technique for neural networks?

Accepted Answer

Dropout regularizes per-forward-pass by randomly zeroing activations; early stopping limits total training iterations — both reduce overfitting via distinct mechanisms

Answer

Dropout and early stopping are identical in mechanism — both limit training steps

Answer

Early stopping adds noise during training while dropout limits the total number of epochs

Answer

Dropout requires a validation set while early stopping does not require any held-out data

Question 29

What is the practical effect of using a very large patience value (e.g., patience=1000) on a small dataset?

Accepted Answer

Early stopping effectively never triggers, allowing the model to overfit without any intervention

Answer

Very large patience causes the learning rate to explode due to accumulated gradient steps

Answer

Large patience always improves model accuracy by allowing more complete convergence

Answer

Patience above 100 is automatically capped by Keras to prevent memory allocation errors

Question 30

Why do practitioners commonly save both a 'latest checkpoint' and a 'best checkpoint' during long training runs?

Accepted Answer

Best checkpoint ensures optimal model quality; latest checkpoint enables resuming training after unexpected interruptions

Answer

The latest checkpoint is used for deployment and the best checkpoint is used only for debugging

Answer

Both checkpoints always contain identical information; saving two is purely hardware redundancy

Answer

Latest checkpoint stores only hyperparameters while best checkpoint stores model weights

Test yourself