Concept
Mixup Training
A data augmentation technique that creates new training examples by blending pairs of existing examples - teaching models to interpolate smoothly between classes rather than making sharp overconfident decisions.
Added May 18, 2026
Standard classification models are trained on clean examples: this image is a cat, that image is a dog, this review is positive, that review is negative. The model learns to map inputs to discrete categories. But the sharpness of these discrete boundaries can lead to overconfidence: models often assign near-100% probability to one class even for ambiguous inputs that a reasonable person would find uncertain.
Mixup is a simple regularisation technique that addresses this by creating blended training examples. Take two training examples and their labels, and create a new example that is a weighted average of both. If image A is a cat (label: [1, 0]) and image B is a dog (label: [0, 1]), a mixup combination at weight 0.3 produces a new training example that is 30% image A plus 70% image B, with label [0.3, 0.7]. The model is trained to produce this blended probability distribution for the blended input.
This sounds strange - you are asking the model to predict a cat-dog hybrid - but the effect on model behaviour is beneficial. Training on blended examples forces the model to maintain calibrated uncertainty between classes, interpolating smoothly rather than making hard decisions. The model learns that the space between classes is not empty - that intermediate examples exist and should receive intermediate predictions.
The results are consistently positive. Models trained with mixup are better calibrated (their confidence scores more accurately reflect true accuracy), more robust to adversarial examples (manipulated inputs designed to fool the model), and often achieve better generalisation on test sets. The improvements are particularly strong on tasks where the boundary between classes is genuinely ambiguous.
Mixup has been extended in many directions. CutMix patches a region from one image into another rather than blending pixel values. FMix creates masks from Fourier space. For language, token-level mixup blends embedding vectors rather than pixel values, and manifold mixup blends at intermediate layers rather than at the input. All share the core intuition: training on interpolations produces models that interpolate gracefully.
Analogy
Teaching a wine taster using blended wines alongside pure varietals, so they develop intuitions about the spectrum between styles rather than just recognising the extremes. A taster who has only learned pure Cabernet and pure Merlot may be confused by a 60/40 blend; one who has trained on blends has learned to read the spectrum. Mixup trains models the same way - on the full spectrum of possibilities, not just the pure cases.
Real-world example
Image classification models trained with mixup on ImageNet consistently show better calibration than those trained without it. When such a model says it is 70% confident an image is a dog, it is correct about 70% of the time - rather than a non-mixup model that might say 95% confident while only being right 70% of the time. This calibration improvement matters significantly in safety-critical applications.
Why it matters
Overconfident models make poor decisions in deployment because their uncertainty estimates cannot be trusted. Mixup is one of the most effective and computationally cheap techniques for producing well-calibrated models. In high-stakes applications where knowing when a model is uncertain is as important as knowing when it is correct, training regularisation techniques like mixup become critical.
In the news
Related concepts