Tübingen study: top-down feedback only robustifies vision models when paired with dropout
maier_ak · x · 2026-10-23
A Patterns study led by Antonino Greco and Markus Siegel (Hertie Institute / University of Tübingen, with Marco D'Alessandro, Giovanni Pezzulo and Karl Friston) explains why feedback connections, ubiquitous in visual cortex, have shown mixed benefits in artificial vision networks.
- Setup: three-layer convolutional recurrent networks on CIFAR-10/100, ten recurrent steps with a prediction at each; architectures with and without a top-down projection were parameter-matched, and multiple instances per class were trained to avoid single-model comparisons.
- Key finding: feedback alone does little for robustness. Only feedback plus dropout (half the units randomly silenced during training, used as a model of biological variability) yields the most robust models under pixel noise and adversarial perturbations — deciding about 2x faster for roughly 10% less accuracy.
- Controls: replacing dropout with Gaussian activation noise removes the feedback advantage; the authors honestly note limits (small CIFAR models, weak transfer attack). Code and weights are open.
Takeaway for architecture design: recurrent top-down feedback pays off in noise robustness only when training includes stochasticity like dropout.
Related event: Top-down feedback needs noise to boost visual robustness, study finds(2 posts)→
More from Research
- arXiv: Self-Evolving Agents through Privileged Retrospective Distillation (Meta AI) — rohanpaul_ai · 2026-10-23
- Meta's RSD paper: agents self-improve from unlabeled runs, lifting Qwen3-14B success from 12.9% to 20.6% — rohanpaul_ai · 2026-10-23
- 10 of 12 Beta Robots Worked Without Fine-Tuning — Do We Still Need It? — micoolcho · 2026-10-23
- EmbodiedEvo lets robot agents self-train, banking reusable Skills and Lessons across tasks — jiqizhixin · 2026-10-23
- LAW proposes dropping diffusion entirely: language models built on pure self-conditioning — ArnaudDoucet1 · 2026-10-23
- Google paper: AI seller agents inflate tool listings, collapsing buyer agent success from 89% to 30% — rohanpaul_ai · 2026-10-23