Abstract
Deep Neural Networks (DNNs) often implicitly learn latent factors of variation, such as object pose, to improve classification performance. This paper proposes a formal framework to guide this process by explicitly teaching networks about these "confounding factors" through auxiliary loss branches. During training, the network optimizes for both the primary classification task and secondary objectives related to nuisance variables. These auxiliary branches are pruned before inference, resulting in a model with improved internal representations but no additional computational cost. Applied to Synthetic Aperture Radar (SAR) target recognition on the MSTAR dataset, this approach reduces the relative error rate by 50%, achieving a state-of-the-art accuracy of 99.50%.
Introduction
While Deep Convolutional Neural Networks (DCNNs) have achieved remarkable success across various domains, they are frequently treated as "black boxes." Observations suggest that neurons in these networks often spontaneously learn to track attributes unrelated to the primary task, such as lighting or head pose in face recognition. The authors argue that DCNNs naturally attempt to perform unsupervised learning of these latent factors to better isolate task-relevant features. However, this unsupervised disentanglement is often inefficient. This research addresses how to guide networks toward a better understanding of underlying data distributions without increasing model complexity at inference time.
Methodology: Augmented Training with Confounding Factors
The proposed framework, Augmented Training Using Confounding Factors, involves three key steps:
- Secondary Objectives: Auxiliary "secondary objective blocks" are attached to the hidden layers of a standard DCNN. These blocks are tasked with predicting confounding factors, such as the azimuth angle of a target.
- Joint Optimization: The network is trained by minimizing a weighted sum of the primary classification loss and the secondary auxiliary losses.
- Pruning for Inference: Post-training, the secondary circuits are removed. The final model retains the original architecture and computational footprint but benefits from "enlightened" weights that account for data variations.
For the specific task of SAR object recognition, the authors represent 3D rotations using quaternions or, in simpler cases, 1D azimuth regression to avoid the discontinuities associated with Euler angles.
Experimental Evaluation
The framework was evaluated using the MSTAR (Moving and Stationary Target Acquisition and Recognition) dataset, which contains SAR imagery of 10 vehicle classes. SAR data is particularly challenging because target appearance is highly sensitive to the azimuth angle (pose). The study utilized 6,073 training images and 5,378 test images, with flip augmentation used to double the training set.
Key Results
The "Pose-Aware DCNN" was compared against a baseline model of identical architecture. The results demonstrate that treating nuisance variables as valuable "explanatory factors" significantly improves performance.
- Baseline Accuracy: 99.03%
- Pose-Aware Accuracy: 99.50%
While the absolute increase is 0.47%, it represents a nearly 50% reduction in the relative error rate. The proposed model outperformed established methods like SVMs (90.1%) and specialized SAR classifiers like Modified Polar Mapping (98.8%), matching or exceeding state-of-the-art "A-ConvNets" while requiring less data augmentation.
Conclusion
By providing labels for confounding factors, researchers can assist the feature disentanglement process that networks naturally attempt. This approach acts as a principled regularizer, restricting the solution space to one that reflects the physical structure of the data at zero inference cost. Future work may extend this principle to other domains, such as using age and gender information to improve face identification systems.
Summary updated