tappiofunktio

tappiofunktio, commonly translated as loss function in English, is a function used in statistics and machine learning to quantify the cost of prediction errors. Given a model's predictions and the true targets, the loss assigns a nonnegative value that typically increases with error. The learning process seeks to minimize the loss over the training data, thereby improving predictive accuracy.

Common examples include mean squared error (MSE) for regression, cross-entropy loss for classification, hinge loss for support vector machines, and L1 or L2 losses for regularization. Robust alternatives such as Huber loss down-weight outliers. The choice of loss depends on the task, data distribution, and desired properties.

Loss functions often possess properties that facilitate optimization. Differentiability enables gradient-based methods; convexity can guarantee a global minimum for certain models. In neural networks, backpropagation computes gradients of the loss with respect to parameters. Regularization terms can be added to the loss, producing a regularized loss function that balances fit and complexity.

In practice, the loss is part of a larger objective function, sometimes called the cost or objective; in probabilistic models, the negative log-likelihood is a common loss. The term tappiofunktio is widely used in Finnish literature to refer to this concept. Python libraries and ML frameworks provide prebuilt loss functions and flexible interfaces to define custom losses.

Notes: The loss value per sample vs. average loss over a dataset; scaling, class imbalance, and outliers influence the choice and interpretation. The loss is not the same as evaluation metrics like accuracy, but it correlates with them during training.