Dataset Augmentation with Generated Novelties

As machine learning models take over an increasingly larger number of domains in our lives, their accuracy, fairness, transparency and adaptability become of greater importance. In the everchanging environments, the resulting ability of the models to perform accurately depends on whether they are able to handle novel, unpredicted and unforeseen instances, examples and classes or any other novel changes in the world of model operation, such as environmental, contextual, distributional changes. The proper handling of novelties sustains the model’s usefulness and adeptness in the long run. The efficiency of response to the encounter of novelties depends on the efforts that were invested at the model training, design and data collection stages. In this work, we propose a variety of approaches and methods which can be incorporated into the novelty generation techniques at the earliest stages of creating the machine learning dataset and the model to assure its robustness and reduce the bias. We revisit distinctions between novelties and anomalies to define a formal novelty generation framework that is domain-agnostic and budget efficient. Then we propose a video-specific use case and evaluate the result of the chosen methods on the video dataset. Our methods aim at making the machine learning solutions adaptable, responsible and show improvement in the accuracy and ability of the models to detect novelties.