Object-Based Sub-Environment Recognition
Keywords: metric learning, environment recognition, bayesian inference, self-supervised learning
TL;DR: The OBSER framework helps real-world agents understand complex environments by recognizing sub-environments using a Bayesian approach with metric learning. Tested in Minecraft and ImageNet, it shows strong generalization and accurate inference.
Abstract: Deep learning agents are advancing beyond laboratory settings into the open and realistic environments driven by developments in AI technologies. Since these environments consist of unique sub-environments, empirical recognition of such sub-environments that form the entire environment is essential. Through sub-environment recognition, the agent can 1) retrieve relevant sub-environments for a query, 2) track changes in its circumstances over time and space, and 3) identify similarities between different sub-environments while solving its tasks. To this end, we propose the Object-Based Sub-Environment Recognition (OBSER) framework, a novel Bayesian framework for measuring object-environment and environment-environment relationships using a feature extractor trained with metric learning. We first design the ($\epsilon,\delta$) Statistically Separable (EDS) function to evaluate to show the robustness of trained representations both theoretically and empirically that the optimized feature extractor can guarantee the precision of the proposed measures. We validate the efficacy of the OBSER framework in open-world and photorealistic environments. The result highlights the strong generalization capability and efficient inference of the proposed framework.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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/2025/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: 6783
Loading