Recommender systems aim to predict personalized rankings based on user preference data.

Thatis,analogousto the zero-shot classification scenario, we nowadvance toannotate novelattributes foradataset via utilizing the knowledge from a few types of seen/given manual attributes, as illustrated in Figure 1. Specifically, we take the well-known CUB dataset [35] as our main test-bed and have a deep investigationonitsattributes.

Machine learning (ML) models are known to leak information about their training data.

34th Conference on Neural Information Processing Systems (NeurIPS 2020), Vancouver, Canada

Existing works have generated explanations for deep neural network decisions toprovide insights into model behavior. Weobservethat these explanations can also be used to identify concepts that caused misclassifications. This allows us to understand the possible limitations of the dataset used to train the model, particularly the under-represented regions in the dataset.

Moreover, CNNs and Transformers struggle to capture global degradation due to their limited receptive fields.