AI 용어집
인공지능 완전 사전
Linear interpolation
Mathematical operation that combines two vectors according to a lambda coefficient between 0 and 1 to generate new intermediate representations.
Mixup Alpha
Parameter of the Beta distribution controlling the degree of mixing in the Mixup technique, where higher values favor more extreme interpolations.
Manifold Mixup
Extension of Mixup that applies interpolation not only to the inputs but also to the intermediate representations of deep neural networks.
AugMix
Mixing technique that combines several chains of augmentative transformations with Mixup to improve robustness to perturbations and spurious correlations.
Remix
Variant of Mixup that resamples the interpolation weights of the labels to correct class imbalance in imbalanced datasets.
Distribution Mixup
Extension of Mixup that directly interpolates the probability distributions of the model's predictions rather than the raw one-hot labels.
Label interpolation
Process of creating hybrid labels by linearly combining the original label vectors proportionally to the mixing of the corresponding inputs.
Implicit regularization
Automatic regularization effect produced by Mixup that constrains the model to behave linearly between training samples.
Interpolation in the feature space
Application of Mixup directly on representations learned by the network rather than on raw inputs for finer control of mixing.
Adaptive Mixup
Variant of Mixup dynamically adjusting the alpha parameter based on sample difficulty or training phase.
Data hybridization
General process of creating new samples by merging existing instances, including Mixup as a particular case of linear combination.
Interpolation vector
Weight vector generated by the Beta distribution determining the mixing proportion between each pair of samples in the Mixup algorithm.
Mixing coefficient
Scalar parameter lambda controlling the intensity of interpolation between two samples, typically drawn from a Beta(α,α) distribution.
Saliency Mixup
Variant guiding the mixing using saliency maps to preserve the most informative regions during sample interpolation.