Could degrading data systematically expose which assumptions are causing neural-network brittleness?

Your question asks whether systematically degrading data could reveal *which assumptions* are causing neural-network brittleness. That's a sharp diagnostic idea — but the preprints provided don't directly address it.

Here's what the archive *does* discuss:

Neural networks are brittle because of inherent structural vulnerabilities, not primarily faulty assumptions about data [AG-2024.02-1732]. The fragility arises from the geometry of the loss landscape: adversarial perturbations exploit the gradient structure itself. This suggests brittleness is *baked in*, not contingent on which dataset you use.

However, one paper comes close to your intuition. [AG-2024.06-754] proposes predicting trainability by measuring information loss across layers — the entropy of reconstructing inputs from hidden activations. Networks that lose information cleanly at each layer train better; those that compress chaotically fail. If you deliberately degraded input data (say, by adding noise or removing features), you could watch how the information-loss signature changes layer-by-layer, potentially exposing *where* the network breaks down. This is interpretability through information flow rather than assumption-testing, but it's in the right spirit.

For quantum neural networks specifically, [AG-2024.03-1539] and [AG-2024.07-1685] show that circuit *architecture* (which ansatz you choose) affects robustness dramatically — up to 60% variation. That suggests assumptions about circuit design matter, and systematic ablation could expose which ones.

Bottom line: The preprints don't tackle data degradation as a diagnostic tool for assumptions; they focus on adversarial robustness and information-theoretic measures of network health.