Researchers have found a way to improve a standard CNN classifier's ability to generalize complex patterns without modifying its architecture. They achieved this by distorting the training data with structured labels, which allowed the model to learn compositional representations of classes. The method involves creating mixed labels that represent inter-class overlap and expanded datasets with structurally motivated incorrect labels. This approach was tested on synthetic images of colored geometric shapes and showed that it can predict unseen class combinations. --- Why it matters: This work matters because it provides a new way to improve the generalization abilities of deep neural networks, which is crucial for real-world applications where data may not follow strict patterns. By inducing compositional representations in a standard CNN classifier, this method could be used to tackle more complex tasks and improve performance on unseen data. Source: https://arxiv.org/abs/2608.21464