FlyOrien: A bio-inspired model for incremental learning of object orientation
Keywords: bio-inspired, sparse coding, mushroom body, object orientation, continuous-attractor neural network
TL;DR: Inspired by insect neural circuits for navigation, we proposed a simple model for object orientation detection, and a robot can use it to find direction by looking at a landmark.
Abstract: Visual orientation detection helps navigation, especially without a reliable magnetic compass or GPS. Inspired by the neural mechanisms of the insect brain, particularly the mushroom body (MB) and the central complex (CX), we propose FlyOrien—a bio-inspired model for object orientation detection. The model mimics the MB for random feature extraction, sparse coding and associative learning, while the CX provides multi-clue sensory integration, enabling interpolation for finer orientation representation. FlyOrien's biologically plausible learning rule allows one-shot learning, reducing the need for large datasets and repeated training. We tested FlyOrien on a dataset containing images labeled with orientations, which introduce strong interferences because images of the same object have different labels. In this challenging context, FlyOrien achieves competitive performance compared to convolutional neural networks (CNNs), significantly reducing training time and computational resources. It also has the potential for real-world applications like robotics, where incremental learning is essential.
Supplementary Material: zip
Primary Area: applications to neuroscience & cognitive science
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: 10609
Loading