Overfitting Calculator

Overfitting Gap is simply Training Accuracy minus Validation Accuracy. A larger positive gap means the model scores much higher on data it was trained on than on unseen validation data, which is the classic symptom of overfitting -- the model has memorized specifics of the training set rather than learning patterns that generalize.

No. The Overfitting Gap is computed directly from the Training Accuracy and Validation Accuracy you enter, with no dependence on Regularization Strength. Regularization Strength only feeds the separate Expected Validation Accuracy output, a projection of what accuracy regularization might recover, not a recalculation of the gap you actually measured.

It applies a common machine-learning rule of thumb -- roughly 10 training samples needed per model parameter for reliable generalization -- by dividing Dataset Size by 10. This is a rough capacity guideline, not a guarantee: the right amount of data per parameter varies substantially by model architecture, task difficulty, and regularization approach.

Variance here is set equal to the Overfitting Gap (high gap = high variance, meaning the model's predictions vary too much between the training set and new data). Bias is estimated as 100 minus Training Accuracy, a proxy for how much the model underfits even its own training data. These are simplified single-number estimates, not a full bias- variance decomposition from repeated resampled training.

Complexity Ratio is Model Complexity (in thousands of parameters) divided by Dataset Size, so raising Model Complexity while holding Dataset Size fixed always raises the ratio proportionally. A higher ratio signals a model with more parameters relative to available training examples, which raises overfitting risk per the same 10-samples-per- parameter guideline used for Optimal Complexity.