A Latent Generative Model for Closed-set and Open-set Recognition
Keywords: open-set recognition, closed-set recognition
Abstract: The classic recognition problem assumes that all possible classes in testing are known in advance during training, which can be termed closed-set recognition (CSR). As a natural extension, open-set recognition (OSR) requires models to reject samples of unknown classes that are not encountered in the training phase. Traditional discriminative models struggle to learn decision boundaries for OSR due to the absence of unknown samples. This has led to existing methods focusing on either CSR or OSR, as optimizing one often results in performance degradation of the other. In this paper, we offer a formalization for OSR based on learning theory, demonstrating that CSR and OSR share the same goal for generative models. Motivated by this core insight, we introduce a neural Latent Gaussian Mixture Model (L-GMM) accompanied by a collaborative training algorithm. The model consists of an encoder that maps inputs to a latent space, and a density estimator that computes probability densities. The end-to-end training algorithm, designed in a collaborative manner, learns the density estimator through maximum likelihood estimation and trains the encoder using a discriminative loss derived from the generative model. This framework yields a model capable of performing both CSR and OSR. Experimental results show that L-GMM outperforms its discriminative counterparts in image recognition and segmentation in CSR with models trained from scratch. These models also outperform other specialized methods when directly applied to OSR without any modifications or prior knowledge.
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4584
Loading