Simple definitional question: In the context of machine learning, is the error of a model always the difference of predictions $f(x) = \hat{y}$ and targets $y$? Or are there also other definitions of error?

I looked into other posts on this, but they are not sufficiently clear. See my comment to the answer in this post:

What's the difference between Error, Risk and Loss?


2 Answers 2


The error can have different forms depending on the application. For example for a simple regression we often use the sum of squared deviations between the actual output $y_n$ for the input $x_n$ and the predicted output $\hat{y}(x_n)$ for the input $x_n$. The total loss $J_\text{Gauss}$ is then given as the sum over all squared errors (also known as Gaussian Loss)for each observation.

$$J_\text{Gauss}= \sum_{n=1}^N\left[y_n-\hat{y}(x_n)\right]^2$$

If we use absolute values instead of squares we obtain the Laplacian loss function $J_\text{Laplace}$, which is given by


If we rather try to compare two probability distributions $p(x)$ and $q(x)$ we use a unsymmetric distance meassure called Kullback-Leibler divergence

$$ {\displaystyle D_{\text{KL}}(P\parallel Q)=\int _{-\infty }^{\infty }p(x)\ln {\frac {p(x)}{q(x)}}\,dx}. $$

For binary classification we can use the hinge-loss

$$J_\text{hinge}=\sum_{n=1}^N\max \{0, 1- t_n \hat{y}(x_n)\},$$

in which $t_n=+1$ if observation $x_n$ is from the postive class and $t_n=-1$ from the negative class.

For support vector regression the $\varepsilon$-insensitive loss $J_\varepsilon$ is used. It is defined by the following equation.


This loss acts like a threshold. It will only count something as an error if the error is larger then $\varepsilon$.

As you can see there are some meassures (see this Wikipedia article for loss functions used for classification) of error for comparing the predicted output and the observed output.

  • $\begingroup$ Hm, but each of the loss functions you describe works on the difference of $y_i$ and $\hat{y}(x_i)$. Even the log loss works on the difference, if you exponentiate it back to non-log probabilities. It seems the error is still just $y_i - \hat{y}(x_i)$ and the different loss functions only do different things with the ever same basic error.... ? Note, that I am using loss and error for different things. The loss is a function of the error. $\endgroup$ Jul 29, 2019 at 13:49
  • $\begingroup$ The Kullback-Leibler divergence does not use errors for different probability distributions. And if you are doing generative modelling you can try to use it for almost all problems. Additionally the hinge loss does not work with a difference of between prediction and true output. $\endgroup$ Jul 30, 2019 at 7:13

Simple definitional question: In the context of machine learning, is the error of a model always the difference of predictions 𝑓(𝑥)=𝑦̂ and targets 𝑦? Or are there also other definitions of error?

No, I don't think the word "error" can be considered as a technical term which always follows this specific definition.

The word "error" is frequently used in the context of evaluation: in a broad sense, an error happens every time a ML system predicts something different than the true value. It's obviously the same concept as in the definition, but it can be applied to non numerical values for which the mathematical difference doesn't make sense.

For example in the context of statistical testing and in classification it's common to talk about "type I errors" and "type II errors" for instances misclassified as false positive and false positive respectively.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.