Deciding When to Switch: E-Processes for Adaptive Minimax Training for Generative Adversarial Nets

This paper addresses the problem of adaptively determining when to switch between discriminator and generator updates during generative adversarial network (GAN) training, replacing fixed update ratios with a statistically principled, data-adaptive scheduling rule. The authors formulate the update-scheduling problem as a sequential hypothesis-testing problem and develop an e-process-based adaptive training procedure.

The paper states: "We formulate this switching problem as sequential hypothesis testing and develop an e-process-based adaptive training procedure. During discriminator updates, one e-process tests the null that the discriminator-induced separation between the empirical data distribution and the generator law remains below a target level. During generator updates, with the discriminator fixed, a second e-process tests the reverse null that this separation remains above a refresh level."

The key methodological contributions are as follows:

1. E-process construction for discriminator updates: The paper considers the composite null hypothesis H0D(aD): TV(P, Q) ≤ aD, where TV is total variation, P is the real data distribution, and Q is the current generator distribution. Two e-variable constructions are proposed based on an arbitrary discriminator D:

• The separable-score e-variable: E sep D(X, Y) = 4σ D(X) σ −D(Y) / (1 + aD)2

• The difference-score e-variable: E diff D(X, Y) = 2σ D(X) − D(Y) / (1 + 2aD − aD2)

Proposition 1 establishes that both are valid e-variables under H0D(aD).

2. E-process construction for generator updates: With the discriminator fixed, the paper considers the operational global null H̃G∞,t(bG): ∆t,l(D̃t) ≥ bG for all l ≥ 0, where ∆t,l measures the discrepancy as witnessed by the fixed discriminator. Proposition 3 provides two e-variable constructions:

• E G,sep(X, Y) = 4σ D̃t(Y) σ −D̃t(X) / (1 − bG)2

• E G,diff(X, Y) = σ D̃t(Y) − D̃t(X) / (1 − bG)

3. Anytime-validity: The paper proves that "Conditional on the observed training sample, we prove that fresh empirical indices and latent draws yield conditional e-values that can be accumulated into e-processes, providing anytime-valid Type I error control under adaptive model updates and data-dependent switching."

4. Oracle e-power characterization: Proposition 2 shows that the logarithmic e-power is maximized when the discriminator approximates the log-density ratio log(p/q), with explicit formulas for the supremum of the expected log-growth for both constructions.

The experimental results demonstrate the effectiveness of the proposed method:

Mixture of 16 Gaussians: The paper reports that "the quality of the generated samples depends strongly on the relative numbers of discriminator and generator updates. The 1D:1G, 5D:1G, and 1D:5G schedules fail to recover the multimodal structure, whereas 5D:5G captures such multimodal structure. The proposed adaptive schedule most faithfully recovers all 16 modes without requiring a prespecified update ratio." This holds across original GAN, hinge, and Wasserstein objectives.

CIFAR-10 benchmark: The paper states: Under both GAN objectives, the adaptive strategy yields the highest mean IS and the lowest mean FID among the training strategies considered. Specifically:

• For softplus loss: Adaptive achieves IS of 8.192 ± 0.080 and FID of 15.877 ± 0.354, outperforming Fixed 1:1 (IS 4.230, FID 83.407), Fixed 5:1 (IS 8.072, FID 17.540), and TTUR (IS 7.139, FID 34.997).

• For hinge loss: Adaptive achieves IS of 8.330 ± 0.079 and FID of 13.208 ± 0.357, outperforming Fixed 1:1 (IS 4.416, FID 80.295), Fixed 5:1 (IS 8.281, FID 13.542), and TTUR (IS 7.508, FID 31.106).

The paper concludes: "We developed an e-process-guided framework for adaptively determining the numbers of discriminator and generator updates in GAN training. Rather than relying on a fixed update ratio, the proposed method uses sequential evidence to decide when to terminate the current optimization phase and switch to the opposing network. The resulting stopping rules are anytime-valid under repeated monitoring. Our construction applies across GAN objectives through variational representations of distributional discrepancies. The framework provides a statistically principled alternative to heuristic update schedules and may extend to other minimax problems requiring sequential assessment of competing components."

Improvements to AI systems:

1. Adaptive training schedulers for generative models: Replace fixed or heuristic update ratios (e.g., 1:1, 5:1) in GANs and other minimax architectures with e-process-based stopping rules that dynamically switch between generator and discriminator updates based on real-time statistical evidence. This eliminates manual hyperparameter tuning of update ratios and improves sample quality (e.g., higher Inception Score, lower FID) across different loss functions (softplus, hinge, Wasserstein).

2. Anytime-valid monitoring for model convergence: Integrate e-processes as continuous, non-stop monitoring tools during training to detect when the discriminator's separation between real and generated distributions falls below a target threshold (or rises above a refresh level). This provides formal Type I error control under repeated checks, enabling safe early stopping or phase switching without risking overfitting to stale discriminator/generator states.

3. Oracle-guided discriminator design: Use the e-power characterization (Proposition 2) to inform discriminator architecture or regularization, encouraging the discriminator to approximate the log-density ratio log(p/q). This improves the efficiency of evidence accumulation, leading to faster and more reliable switching decisions, and can be applied to any adversarial training where a critic estimates a distributional discrepancy.

4. Generalizable minimax optimization: Extend the framework to other two-player adversarial systems (e.g., adversarial robustness training, actor-critic reinforcement learning, or federated learning with generator-discriminator dynamics) where sequential assessment of competing components is needed. The e-process approach provides a principled, data-adaptive way to allocate computation between optimizing one component versus the other, reducing wasted iterations and improving final performance.

5. Self-tuning generative pipelines: Build AI systems that automatically adjust their own training schedules in real time, without human intervention, by embedding e-process-based switching into standard training loops. This improves robustness across datasets and objectives, as demonstrated by consistent gains on CIFAR-10 and multimodal synthetic data, and reduces the need for extensive hyperparameter searches.

What the improved AI system can do:

• Train GANs (or similar minimax models) with no fixed update ratio, automatically balancing discriminator and generator updates to achieve better mode coverage and higher fidelity (e.g., recovering all 16 Gaussian modes where fixed ratios fail).

• Monitor training progress with statistically valid, anytime-valid tests, allowing safe early termination or phase switching without invalidating results due to repeated peeking.

• Adapt to different loss functions and data distributions without re-tuning, as the e-process construction is objective-agnostic via variational representations.

• Provide interpretable evidence (e-process values) for why a switch is made, aiding debugging and transparency in generative model training.

• Extend to other adversarial or competitive AI settings (e.g., robust classifiers, multi-agent RL) to allocate optimization effort dynamically, improving convergence speed and final performance.