Model Evaluation & Optimization Techniques Flashcards

Which regularization technique explicitly prevents co-adaptation of neurons by randomly zeroing activations during training?

Answer: Dropout

Dropout randomly deactivates neurons during each training step, forcing the network to learn redundant representations and preventing co-adaptation.

In the context of learning curves, what does a large gap between training and validation loss with both curves converging indicate?

Answer: High variance (overfitting) needing more regularization or data

A large persistent gap between training and validation loss with convergence indicates the model has memorized training data and is not generalizing well (high variance/overfitting).

What does the Brier Score measure in probabilistic classification?

Answer: The mean squared error between predicted probabilities and actual binary outcomes

The Brier Score is the mean squared error between predicted class probabilities and the true binary (0/1) outcomes, rewarding well-calibrated probability estimates.

Which cross-validation strategy is most appropriate when time-series data exhibits temporal dependencies?

Answer: Time series split with expanding window

Time series split uses an expanding or sliding window where the validation set always comes after the training set chronologically, preserving temporal causality.

What is the effect of using a very small batch size in stochastic gradient descent?

Answer: Higher gradient noise that can escape local minima but slower wall-clock convergence

Small batch sizes produce noisy gradient estimates that can help escape sharp local minima but typically require more iterations and wall-clock time to converge.

In multi-objective hyperparameter optimization, what does the Pareto frontier represent?

Answer: The set of configurations where no objective can be improved without degrading another

The Pareto frontier contains all non-dominated solutions — configurations where improving one objective (e.g., accuracy) necessarily worsens another (e.g., inference latency).

Which technique decomposes model predictions into feature contributions that satisfy both efficiency and local accuracy properties inspired by cooperative game theory?

Answer: SHAP (SHapley Additive exPlanations)

SHAP uses Shapley values from cooperative game theory to fairly attribute each feature's contribution, satisfying efficiency (contributions sum to prediction) and local accuracy.