DNNs Generalize to Novel Object Orientations via Shared Feature Neurons
A recent study published on arXiv (paper 2109.13445) reveals that Deep Neural Networks (DNNs) can identify objects in unfamiliar orientations by utilizing orientation-invariance derived from previously encountered objects at various angles. This research, which falls under the Computer Vision and Pattern Recognition category, demonstrates that the ability to generalize enhances with an increase in familiar objects, particularly regarding 2D rotations. It implies that neurons activate in response to shared characteristics between known and unknown objects, reflecting processes similar to those in the brain. This skill emerges without the need for specific training on new orientations, presenting a significant property of deep learning. The results have repercussions for both computer vision and neuroscience, questioning earlier beliefs about DNN constraints. The study, along with its code and data, is available to the public.
Key facts
- DNNs can generalize to objects in novel orientations by disseminating orientation-invariance from familiar objects.
- Generalization improves with an increasing number of familiar objects.
- Generalization only occurs for 2D rotations of familiar orientations.
- Neurons tuned to common features between familiar and unfamiliar objects enable this dissemination.
- The study is published on arXiv under paper number 2109.13445.
- The paper is categorized under Computer Science > Computer Vision and Pattern Recognition.
- The research suggests brain-like neural mechanisms for generalization.
- The study was submitted to arXiv and is available with code and data.
Entities
Institutions
- arXiv