Loss as Training Signal
Loss is a signal for training, not the whole goal.
A lower loss usually means the model is better under the chosen loss function. But it does not automatically mean the model is useful, fair, robust, or correct for the real task.
The loss answers a narrower question:
How bad was this prediction according to this rule?
Training uses that number to adjust parameters later.
Match loss to task
Different tasks need different losses.
regression -> mean squared error is a common starting point
binary classification -> binary cross-entropy
multiclass classification -> multiclass cross-entropy
The loss should match what the output means.
If the model predicts one continuous number, MSE may be appropriate. If it predicts a probability for class 1, binary cross-entropy fits. If it predicts one class among many, multiclass cross-entropy fits.
A model predicts one continuous number, such as a temperature. Enter 1 for MSE or 0 for multiclass cross-entropy.
Compute it first, then check your number.
HintContinuous target
Regression predicts numbers, not class probabilities.
SolutionWork it out
A continuous-number prediction is a regression task. Mean squared error is a common starting loss for regression.
A classifier chooses one of 5 classes. Enter 1 for multiclass cross-entropy or 0 for mean squared error as the usual first choice.
Compute it first, then check your number.
HintOne class among many
The model should assign probability to the true class.
SolutionWork it out
For one-of-many classification, multiclass cross-entropy is the usual first choice because it rewards high probability on the true class.