Discovering Novelty via Transfer Learning

Modern-day intelligent systems, i.e., autonomous vehicles, smart manufacturing, etc., rely on distinct data-driven machine learning (ML) models for performing different safety-critical operations such as pedestrian detection, chemical plant control, among others. Since the performance of ML models depends on their generalizability capability, handling out-of-distribution data during the operational phase is of paramount importance for enhancing the adaptability of artificial intelligence (AI) systems. Hence, finding the root causes of novelty is critical for minimizing its impact on the performance of AI systems. However, detection of novelties is not trivial since each AI system operates in a specific environment or agent settings, whereas small changes cause the detection mechanism to adapt to different environmental constraints. For example, detecting novelties in identical intelligent navigation systems differs if the system is deployed in two different countries, though the operation of a navigation system is indistinguishable. In this paper, to reduce the dependability of novelty detectionmechanisms on environment's or agent's attributes, we propose transfer learning-based novelty detection mechanisms for inter-domain applications. In this regard, we analyze the importance of feature transformation to enhance novelty detection systems' transferability. The proposed detection mechanisms aim at augmenting AI systems with rapid responsiveness to novel surroundings, thus, making AI systems responsible and trustworthy. We conduct multiple experiments on state-of-the-art neural networks, i.e., ResNet50, Mobile-Net, and benchmark datasets, i.e., MNIST.