NOD-CC: A Hybrid CBR-CNN Architecture for Novel Object Discovery JT Turner 1 , Michael W. Floyd 2 , Kalyan Gupta 2 , and Tim Oates 1 1 University of Maryland Baltimore County, Baltimore, MD, USA jturner1@umbc.edu, oates@umbc.edu 2 Knexus Research Corporation, National Harbor, MD, USA first.last @knexusresearch.com Abstract. Deep Learning methods have shown a rapid increase in pop- ularity due to their state-of-the-art performance on many machine learn- ing tasks. However, these methods often rely on extremely large datasets to accurately train the underlying machine learning models. For super- vised learning techniques, the human effort required to acquire, encode, and label a sufficiently large dataset may add such a high cost that de- ploying the algorithms is infeasible. Even if a sufficient workforce exists to create such a dataset, the human annotators may differ in the quality, consistency, and level of granularity of their labels. Any impact this has on the overall dataset quality will ultimately impact the potential perfor- mance of an algorithm trained on it. This paper partially addresses this issue by providing an approach, called NOD-CC, for discovering novel object types in images using a combination of Convolutional Neural Net- works (CNNs) and Case-Based Reasoning (CBR). The CNN component labels instances of known object types while deferring to the CBR com- ponent to identify and label novel, or poorly understood, object types. Thus, our approach leverages the state-of-the-art performance of CNNs in situations where sufficient high-quality training data exists, while min- imizing its limitations in data-poor situations. We empirically evaluate our approach on a popular computer vision dataset and show significant improvements to object classification performance when full knowledge of potential class labels is not known in advance. Keywords: Deep Learning · Novel Object Discovery · Computer Vision · Convolutional Neural Networks. 1 Introduction Deep Learning has seen rapid advancement in recent years, setting benchmarks for many machine learning tasks in the areas of computer vision, natural lan- guage processing, and game AI. While these deep neural networks are fundamen- tally the same as the perceptrons [14] of the late 1960s, they leverage dramatic improvements in the availability of computational resources and training data to significantly outperform their predecessors. In particular, the field of com- puter vision has benefited from the application of Convolutional Neural Net- works (CNNs) [6] that are able to use massive image datasets to learn relevant