A new probe reads any trained network as a stack of input-dependent filters — and finds a top denoiser has been quietly re-deriving a classical 1990s smoother.

Deep learning's most-used denoising networks may have been quietly re-deriving classical image filters for years. The reason this stayed invisible is that standard explanation tools, gradient-based saliency and perturbation analysis, only work on differentiable networks and treat the model as a black box, when the model is often doing something much more legible: local, input-dependent filtering.

An arXiv preprint's finding that the DnCNN denoiser weights pixels by spatial and gray-value distance, in exactly the pattern classical bilateral filtering uses, is the cleanest demonstration yet. The probe the paper introduces, called neural echoes, treats any trained network as a stack of input-dependent filter banks and reads them directly, without requiring gradient access. That matters because the same method works on transformers, recurrent models, and quantized networks where the standard tools fail.

The reusable mechanism: when an explanation method is gradient-bound, you are reading the network's loss surface, not its behavior. Read the filters instead. The pattern here is not that one network copied a classical filter. It is that explainability has been blocked by tool choice, not by the networks themselves.