Probing the Limits: Adversarial and OOD Robustness of Neural Thicket Ensembles

TL;DR: If we build ensembles from random perturbations around a pretrained model, do they stand up to tricky attacks or weird new data? We’ll stress-test these ensembles and seek ways to make them even more robust.

Research Question: How robust are neural thicket ensembles to adversarial attacks and out-of-distribution (OOD) data, and can their resilience be enhanced by targeted diversity-promoting perturbations?

Hypothesis: While neural thicket ensembles improve robustness compared to single models, their vulnerability to adversarial and OOD inputs depends on the diversity and independence of the sampled experts. Explicitly encouraging diversity (e.g., via orthogonal perturbations or adversarially selected directions) will further enhance ensemble resilience.

Experiment Plan: Construct ensembles using random perturbations, orthogonal perturbations, and adversarially chosen perturbation directions around pretrained weights. Evaluate on standard adversarial attack benchmarks and OOD datasets, comparing ensemble types. Metrics: adversarial accuracy, OOD detection rates, ensemble diversity. Expected outcome: Diversity-aware neural thicket ensembles outperform vanilla ensembles in both robustness and generalization.

References:

• Gan, Y., & Isola, P. (2026). Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights.
• Cai, Y., Ning, X., Yang, H., & Wang, Y. (2023). Ensemble-in-One: Ensemble Learning within Random Gated Networks for Enhanced Adversarial Robustness. AAAI Conference on Artificial Intelligence.