Abstract
Traditional neural network architectures often rely on cooperative subnetworks working toward a single objective. In this work, we propose "Advocacy Learning," a competitive framework for classification where class-specific "Advocates" compete to convince a "Judge" of their respective class hypotheses. By training Advocates to be "deceptive"—maximizing the Judge's confidence in their class regardless of the true label—the system learns robust, class-conditional representations. Our experiments on MNIST, Fashion-MNIST, and CIFAR-10 demonstrate that this adversarial pressure improves discriminative performance and robustness to data imbalance compared to standard attention-based models.
Introduction
Most deep learning attention mechanisms are saliency-based, highlighting regions generally important for classification but often failing to provide class-specific evidence. They tell the model where to look, but not necessarily what evidence supports a specific hypothesis. Furthermore, standard training is purely cooperative. Drawing inspiration from adversarial frameworks like GANs and competitive systems in law or economics, we investigate whether a competitive environment can uncover more potent discriminative features. We introduce the Advocacy Net, where subnetworks are incentivized to "argue" for their specific class, forcing a central Judge to become highly discerning.
Methods
The Advocacy Net architecture consists of two primary components: a panel of $N$ Advocates and a single Judge.
- Advocates: For every possible class, an Advocate produces a class-conditional attention map, or "argument" ($a^i$). This map highlights features in the input $x$ that support its assigned class.
- The Judge ($J$): The Judge receives the "evidence" ($e_i = x \odot a^i$) from all $N$ Advocates and predicts the true class $\hat{y}$.
The core innovation lies in the "deceptive" training objective. While the Judge minimizes standard cross-entropy to ensure accuracy, each Advocate $i$ is trained to maximize the Judge's confidence in class $i$, even when the input belongs to a different category. This adversarial pressure forces Advocates to find the most convincing class-specific evidence and requires the Judge to learn subtle boundaries to avoid being misled.
Results and Discussion
We evaluated Advocacy Learning across several image datasets (MNIST, Fashion-MNIST, CIFAR-10) and a medical time-series dataset (MIMIC-III).
Performance Gains
On Fashion-MNIST, Advocacy Learning achieved a 91.62% accuracy, representing a 19% relative reduction in error compared to a standard attention baseline (89.71%). The model also showed significant robustness to class imbalance, with a 37% error reduction in skewed MNIST variants.
Key Insights
- Deception as a Catalyst: Models trained with "deceptive" advocates outperformed those where advocates were only trained on their own class ("Honest Advocacy"). This confirms that competition drives the discovery of more effective features.
- Capacity Constraints: Success requires a capacity imbalance; the Judge must be sufficiently complex to resist simple "noise-based" deception from the Advocates. If Advocates are too powerful, the system collapses into nonsensical arguments.
- Domain Limitations: The framework was less effective on 1D physiological signals (MIMIC-III). Unlike images, where pixels can be masked to mimic other objects, 1D signals lack the morphological flexibility required for effective "argumentation" via masking.
Conclusion
Advocacy Learning demonstrates that competition between subnetworks can lead to superior class-conditional representations. While currently limited by scalability—requiring one subnetwork per class—future work into hierarchical advocacy could extend this approach to large-scale datasets. The resulting class-conditional maps also offer a promising path toward more interpretable deep learning models, providing a clear window into the evidence used for specific categorical decisions.
Summary updated