Słownik AI
Kompletny słownik sztucznej inteligencji
Validation curve
Graph showing model performance on a separate validation set alongside the learning curve to detect overfitting.
Convergence point
Moment when the learning and validation curves stabilize, indicating that adding more data will no longer significantly improve performance.
Generalization gap
Difference between performance on training data and validation data, measuring the model's ability to generalize to new data.
Training error
Measure of the error made by the model on the training set, serving as a reference to evaluate potential overfitting.
Validation error
Model performance evaluated on a dataset not used during training, reflecting its actual generalization capability.
Learning plateau
Phase where adding additional training data no longer produces significant improvement in model performance.
Bias diagnosis
Analysis of the learning curve to determine if the model suffers from underfitting (high bias) requiring a more complex model.
Variance diagnosis
Identification of overfitting (high variance) when training error is low but validation error remains high.
Model complexity
Factor influencing the shape of the learning curve, where an overly complex model tends to overfit while an overly simple model underfits.
Critical sample size
Minimum number of training data points required to achieve stable and generalizable model performance.
Per-class learning curve
Separate analysis of learning curves for each output class, particularly useful in imbalanced classification problems.
Marginal improvement rate
Measure of the performance gain obtained by adding additional units of training data, helping to decide the relevance of collecting more data.
Iterative cross-validation
Technique combining cross-validation and learning curves to robustly evaluate performance at different sample sizes.
Noise in data
Random error in training data that affects the shape of the learning curve and limits the achievable performance of the model.