← AI Terminology

MSE - Mean Squared Error

Mean Squared Error (MSE) is a regression loss function and evaluation metric that measures the average squared difference between predictions and actual values — disproportionately penalising large errors due to squaring.

It is the most common loss function for regression tasks in machine learning.
Why It Matters in AI
MSE is mathematically convenient (differentiable everywhere, smooth gradient), statistically principled (maximum likelihood estimator under Gaussian noise), and heavily penalises large errors — making it appropriate when large mistakes are disproportionately costly. It is the default regression loss in almost all ML frameworks. However, its sensitivity to outliers means a single bad prediction can dominate the metric — understanding this tradeoff vs. MAE is essential for metric selection.
Key Points
Aspect Description
RMSE Root Mean Squared Error — √MSE, same units as target, preserves outlier sensitivity of MSE
Units Squared units of target — less interpretable than MAE (use RMSE = √MSE for interpretability)
vs MAE MSE penalises large errors more heavily (squares them); MAE is robust to outliers
Formula MSE = (1/n) × Σ (yᵢ − ŷᵢ)² — mean of squared differences between predictions and actuals
L2 loss MSE is equivalent to L2 loss used as a training objective
Gradient ∂MSE/∂ŷᵢ = 2(ŷᵢ − yᵢ) — smooth, proportional to error magnitude — ideal for gradient descent
Simple Analogy
A judge scoring diving: a diver who splashes a lot (big error) gets penalised heavily because the score is based on the square of the splash size. Small splashes (small errors) barely affect the score. MSE builds this "big mistakes matter disproportionately" rule into every regression training run.
Common Usage Examples
  • sklearn.metrics.mean_squared_error(y_true, y_pred) — evaluation
  • loss = nn.MSELoss()(predictions, targets) — PyTorch regression training loss
  • rmse = np.sqrt(mean_squared_error(y_true, y_pred)) — RMSE for interpretable reporting
  • Image super-resolution: pixel-level MSE between predicted and ground-truth high-res images
  • keras.losses.MeanSquaredError() — Keras regression loss
Summary
In short: MSE is the standard regression loss — differentiable, statistically principled, and heavily penalising large errors — making it the default choice for training regression models and evaluating continuous predictions.