Abstract
Standard Convolutional Neural Networks (CNNs) struggle with rotation invariance, often requiring massive data or parameters to recognize objects at arbitrary angles. This paper introduces CRGEN (Continuous Rotation Group Equivariant Network), a biologically inspired framework that leverages bell-shaped "tuning curves" from neural population coding. By modulating convolution kernels with Gaussian functions and using learnable sparse weights, CRGEN achieves continuous rotation symmetry. Experimental results demonstrate state-of-the-art performance on rotated image benchmarks while utilizing up to 75% fewer parameters than previous methods.
Introduction
While CNNs are inherently translation-invariant, they lack robustness to rotation. A model trained on upright images often fails to recognize the same objects when rotated. Existing Group Equivariant CNNs (G-CNNs) typically address this through discrete rotations (e.g., 90-degree steps), but achieving continuous rotation equivariance usually entails high computational costs or complex mathematical approximations. This work looks to neuroscience—specifically how the brain encodes orientation—to bridge the gap between discrete digital grids and continuous physical rotations.
Methodology: Bio-inspired Tuning Curves
The core innovation of CRGEN is inspired by Neural Population Coding. In biological systems, neurons respond to stimuli based on "tuning curves"—bell-shaped response functions centered on a preferred orientation.
The authors implement this through several key mechanisms:
- Gaussian Modulation: Convolution kernels are modulated with Gaussian functions, creating overlapping receptive fields in the orientation dimension. This allows the network to "interpolate" between discrete angles.
- Sparse Weights with Learnable Geometry: Instead of dense pixel grids, kernels are generated from sparse points whose locations and orientations are learnable parameters.
- Equivariance in Continuous Space: By ensuring tuning curves are sufficiently wide (overlapping), the network achieves near-perfect continuous equivariance even on discrete grids.
To maintain this symmetry in deep architectures without "blurring" features, the authors also introduced CRGEN-LN, which utilizes an equivariant loss to preserve symmetry across multiple layers.
Experimental Results
The model was evaluated on benchmarks including MNIST-rot, CIFAR-10, and ImageNet-100.
Efficiency and Parameter Reduction
On the MNIST-rot dataset, CRGEN-S achieved a 0.85% error rate, surpassing the previous state-of-the-art (SFCNN, 0.88%). Remarkably, it did so using only 40K parameters, a 160x reduction compared to SFCNN’s 6.5M parameters.
Small-Sample and Zero-Shot Learning
Equivariant networks possess a smaller hypothesis space, making them highly effective when data is scarce. When trained on only 200 samples, CRGEN achieved 80.17% accuracy, significantly outperforming standard CNNs (<50%). In "zero-shot" tests (training on upright, testing on rotated), CRGEN maintained over 90% accuracy across all angles, whereas standard CNN performance dropped below 20% for inverted images.
Scaling to Real-World Data
Using a ResNet-18 backbone on ImageNet-100, CRGEN-LN outperformed standard G-CNNs by over 21%, demonstrating its scalability to high-resolution, complex natural images.
Discussion and Future Work
CRGEN demonstrates that smoothness is the key to continuous symmetry. By ensuring overlapping receptive fields, networks can achieve continuous rotation invariance without continuous grids. However, challenges remain:
- Computational Speed: Sparse convolutions currently lack the highly optimized GPU implementations available for standard dense convolutions.
- Resolution Loss: Stacking many smooth layers can lead to blurring; while CRGEN-LN mitigates this, architectural solutions are still being explored.
- Broader Symmetries: Future work may extend this framework to other groups, such as scaling or 3D rotations ($SO(3)$).
In summary, CRGEN provides a parameter-efficient, biologically grounded path toward truly rotation-invariant deep learning.
Summary updated