1. Apply process to your dev and test sets to make sure they come from the same distribution
2. examining examples your algorithm got right as well as the ones it got wrong
3. now train and dev/test set data may come from slightly varying distributions from one another

University of Michigan - Ann Arbor

A test dataset is a separate partition of data that is held out from the learning process and used strictly to evaluate a model's performance. Because performing well on the training data does not guarantee success on unseen data, the test set serves as a critical measure of the model's true ability to generalize.

Test Dataset

How to correct incorrect dev/test set data

Overfitting occurs when a model performs significantly better on its training data than on validation or holdout data, resulting in a large generalization gap. This indicates that the model is overly complex and has begun to memorize noise rather than learning generalizable patterns. However, severe overfitting is not always inherently detrimental in deep learning; the best predictive models often exhibit this behavior. The primary goal is to drive the overall generalization error lower, meaning the gap is only problematic if it obstructs that goal.

Overfitting

For a given classification model $$f$$ and a finite dataset $$\mathcal{D}$$ containing $$n$$ examples, the empirical error, denoted as $$\epsilon_\mathcal{D}(f)$$, is the fraction of instances where the model's prediction disagrees with the true label. It is defined mathematically using the indicator function $$\mathbf{1}$$ as:
$$\epsilon_\mathcal{D}(f) = \frac{1}{n}\sum_{i=1}^n \mathbf{1}(f(\mathbf{x}^{(i)}) 
eq y^{(i)})$$

Empirical Error of a Classifier

When evaluating multiple classifiers, such as $$f_1, \dots, f_k$$, on the same test dataset, the probability of obtaining a misleading test set performance score for at least one model increases significantly compared to evaluating a single classifier. For a single classifier $$f$$, a practitioner might be highly confident that its empirical test error $$\epsilon_\mathcal{D}(f)$$ is close to its true population error $$\epsilon(f)$$. However, as the number of classifiers $$k$$ grows, the risk of a false discovery compounds, making it difficult to guarantee that the best-performing model did not simply achieve its seemingly low error rate by chance. This phenomenon directly relates to the statistical challenge of multiple hypothesis testing.

Multiple Hypothesis Testing in Model Evaluation

In machine learning, the mathematical guarantees of test set performance rely on the assumption that a classifier is chosen without any prior contact with the test dataset. If a subsequent model $$f_2$$ is designed after the modeler observes the test set performance of a prior model $$f_1$$, information from the test dataset inevitably leaks to the modeler. Because of this leakage, the test dataset can no longer be viewed as being drawn randomly from the underlying population relative to the modeler's choices. This human-in-the-loop bias is known as adaptive overfitting and compromises the validity of subsequent test set evaluations.

Learn Before

Related