Neighbourhood generation

Neighbourhood generation involves algorithms that systematically create variations of an original data sample. These variations are designed to be 'close' to the original in feature space, effectively defining a neighborhood around it. Several methods exist for generating these neighbors, including adding noise, applying small transformations (e.g., rotations or translations for images), or using generative models conditioned on the original sample. In the context of adversarial robustness, neighborhood generation helps to identify vulnerabilities in models by exposing them to slightly perturbed inputs. In data augmentation, it expands the training dataset, making the model less sensitive to specific data instances and more capable of generalizing to unseen data. The specific method for neighbourhood generation depends on the data type and the problem at hand; for example, in image recognition, slight rotations or color adjustments might be used, while in natural language processing, synonym replacement or back-translation could be employed. The size and diversity of the generated neighborhood are crucial parameters that affect the effectiveness of the approach; a well-chosen neighbourhood should be large enough to capture relevant variations but not so large as to overwhelm the model with irrelevant data.