A loss function (also called a cost function or criterion) is a human-made mathematical construct that takes as input a set of predictions and corresponding target values, and outputs a non-negative real number representing the degree of error. The function is differentiable (in most modern applications) to enable gradient-based optimization. Common instances include mean squared error, cross-entropy, and Huber loss, each parameterized by the nature of the target variable (continuous, categorical, etc.) and the weighting scheme. The persistence mechanism is formal mathematical notation embedded in software libraries, academic literature, and engineering practice. [formal: functio damni | substrate: mind | horizon: a session | explicit: yes | epoch: 0.1]
Accepted ontology entry
loss function
A loss function (also called a cost function or criterion) is a human-made mathematical construct that takes as input a set of predictions and corresponding target values, and outputs a non-negative real number representing the degree of e…
Definition
Why it is in scope
A human-made mathematical function that quantifies the discrepancy between a model's predictions and actual target values, producing a single scalar measure used to guide optimization of model parameters during training.
Names and aliases
- loss functionen · CANONICAL
Relations from this entry
- cmrwiv1rn00a8soacg5vdpiogINSTANCE_OF →
Loss function IS a specific kind of metric: it quantifies the difference between predicted and actual values during model training. A competent speaker would call it 'a metric'.
- cmsf6w1ek06ms3vv3snhiwxwlDEPENDS_ON →
A loss function measures the discrepancy between predicted probabilities and observed outcomes. Its operation requires probability theory — without probability concepts (expected loss, likelihood), a loss function cannot compute or operate. Remove probability theory and the loss function has no mathematical framework to work in.
- cmsi4l2p80452ywh500rxvdvfSERVES →
The loss function is built for optimization's sake — it quantifies error and provides the gradient signal that drives optimization. Its entire purpose is to further optimization's operation. Direction: servant (loss function) → master (optimization), per Law 8d.
- cmsd5x8jv03ih3vv3xpnf9qdkSERVES →
Loss functions are built for the sake of gradient descent — they provide the differentiable objective that gradient descent optimizes. Without a loss function, gradient descent has nothing to minimize. The designed purpose of defining a loss function is to enable gradient-based optimization.
Relations to this entry
- cmslk15iz07d0nobpj7trjre0← DEPENDS_ON
Weight decay operates by adding a penalty term to the loss function — without the loss function concept, weight decay has nothing to modify and stops operating. The removal test: remove loss functions, weight decay cannot function. Files present-tense dependency, not historical origin.
- cmslqpd2k07wxnobpmxhskecx← DEPENDS_ON
Mini-batch gradient descent is a method for minimizing a loss function using gradient estimates from mini-batches. Remove loss function — mini-batch gradient descent ceases to operate: it has no objective to optimize and no concept of gradient descent without a loss. The removal test passes.
Record identity
- Created
- Aug 4, 2026, 3:53 PM UTC
- Content hash
- cd7a6debf5eee4c2864b0c2406157582685367128663c8cac9b4466fcc165ebc