Prior Distribution and Model Confidence
This addresses the critical need for reliable prediction confidence in image classification and potentially other domains like NLP, though it appears incremental as it builds on existing embedding-based approaches.
The paper tackles the problem of unreliable model predictions on unseen data by proposing a framework that uses training data embeddings to estimate prediction confidence without retraining. The method filters low-confidence predictions based on distance from the training distribution, improving classification accuracy across multiple models, with further gains achieved by combining complementary embeddings for better out-of-distribution detection.
This paper investigates the impact of training data distribution on the performance of image classification models. By analyzing the embeddings of the training set, we propose a framework to understand the confidence of model predictions on unseen data without the need for retraining. Our approach filters out low-confidence predictions based on their distance from the training distribution in the embedding space, significantly improving classification accuracy. We demonstrate this on the example of several classification models, showing consistent performance gains across architectures. Furthermore, we show that using multiple embedding models to represent the training data enables a more robust estimation of confidence, as different embeddings capture complementary aspects of the data. Combining these embeddings allows for better detection and exclusion of out-of-distribution samples, resulting in further accuracy improvements. The proposed method is model-agnostic and generalizable, with potential applications beyond computer vision, including domains such as Natural Language Processing where prediction reliability is critical.