Concept Maps That Cannot See the Noise: Cross-View Consistency for Spatial Concept Bottlenecks on Real Sensor Data

Some image classifiers are built to explain themselves. Before predicting, the network draws a set of "concept maps," heatmaps saying where in the image it sees, for instance, a striped wing or a speckled breast. It then makes its decision from those maps alone. On clean photographs this design now costs almost no accuracy. Real images from cameras and microscopes, however, carry noise that is correlated across neighbouring pixels. Blotches from demosaicing and compression, and grainy clumps in fluorescence images, look like texture. We argue that concept maps trained and tested on such data will partly light up on the noise. The explanation then changes when the same scene is photographed twice, even though nothing real changed. No one has measured this. The standard tests of explanation quality perturb inputs with white noise or gradient attacks, never with the correlated noise real sensors produce.

The proposal has two parts. The first is a measurement: we use repeated captures of the same physical scene to test whether a concept map is driven by the scene or by the noise realization. The second is a training change: we split each image into pixel-shuffled sub-images whose noise is nearly independent, and require the concept maps computed on different sub-images to agree. The reasoning mirrors Noise2Noise-style denoising. Signal is shared across the views and noise is not, so the maps that best agree across views cannot depend on the noise. The claim to test is that this consistency makes concept maps stable across noise realizations and stops noise from leaking into a concept's evidence. We expect it to do so without the accuracy loss that pruning causes, and better than simply denoising the image first.

What the literature establishes and what it leaves open

Spatial concept bottlenecks supply the architecture we build on. Show and Tell (SALF-CBM) keeps concept activations as 2D maps rather than pooled scores. This removes the usual accuracy penalty: it reports 75.32% ImageNet top-1 against 71.95% for the prior label-free concept model, and 76.94% zero-shot segmentation pixel accuracy on CUB. All of its evaluation is on clean imagery.

Two negative results constrain the design. Do Concept Bottleneck Models Obey Locality? shows that a concept's prediction changes when pixels outside that concept's region are perturbed, and that this leakage grows with model capacity. Pruning reduces the leakage but costs task accuracy, so shrinking the model is not an acceptable fix. Its perturbations are also synthetic and applied to clean data.

Denoising supplies both the mechanism and the main competitor. AP-BSN explains how self-supervised denoisers fail on correlated noise: they copy neighbouring noise into the prediction instead of removing it. It restores the needed independence with pixel-shuffle downsampling at a training stride large enough to decorrelate the noise. FM2S models spatially correlated mixed Poisson-Gaussian noise in fluorescence microscopy with a 3.5k-parameter network. Both methods target pixel fidelity and say nothing about what a downstream classifier attends to.

Explanation-metric work supplies the evaluation tools and one gap. Maeß et al. score explanation stability by the largest singular value of the linear map from outputs to input attributions. This implicitly assumes isotropic noise and is applied only to post-hoc heatmaps. CF² defines necessity and sufficiency of an explanation by removing it or keeping only it, but only for graphs. ICE motivates making concept maps non-negative. It finds that non-negative factorizations of activations give additive, part-like concepts that humans prefer, while signed factorizations reconstruct better but are less readable.

Our proposal differs from each of these. SALF-CBM and the locality paper never introduce physical noise. AP-BSN and FM2S never attach a concept layer. Maeß et al. never analyse concept bottlenecks or correlated perturbations.

Core Claim

Pixel-shuffle views are the right lever at the concept level, even though they are awkward at the pixel level. AP-BSN needs asymmetric strides because downsampling destroys the high-frequency detail that pixel restoration must reproduce. Concept maps live at feature-map resolution, typically 14×14 or coarser. A decision about a coarse region survives stride-2 to stride-4 subsampling of a high-resolution input far better than a pixel-exact reconstruction does.

We therefore train a spatial concept bottleneck with three terms. The first is the usual task loss. The second is a cross-view agreement loss between concept maps computed from different stride-$s$ sub-images of the same noisy image, using a stop-gradient teacher so the maps cannot collapse. The third, a non-negativity constraint (softplus), makes each concept map an additive evidence field, so that masking a region removes only positive evidence and necessity and sufficiency tests are well defined.

The stride $s$ is set from the measured noise autocorrelation width, the step Bordone Molini et al. take for blind-spot size. Training needs only noisy data, which matters in microscopy, where clean references are expensive.

Research Program

The measurement comes first, because it must work regardless of whether the method does.

We define noise-realization concept variance. For a scene imaged repeatedly, it is the variance of each concept map across independent captures, divided by the variance of that map across different scenes. It is a signal-to-noise ratio for explanations. The FMD dataset records 50 frames per field of view on confocal, two-photon and wide-field microscopes, and SIDD records long frame sequences per scene. Both supply real repeated noise without simulation.

We also define noise leakage, a physically grounded version of the locality paper's test. We resample only the noise outside a concept's annotated region, keep everything inside fixed, and record how much that concept's in-region activation moves.

Finally, we make the Maeß stability score noise-aware. Let $J$ be the Jacobian of the concept maps with respect to the input, and $\Sigma$ the empirical noise covariance estimated from repeated captures. The worst-case amplification of real noise is then $$\sigma_1\!\left(J\,\Sigma^{1/2}\right),$$ the largest singular value of the Jacobian after whitening by the noise. Under white noise it reduces to the original quantity. We compute it with power iteration rather than full SVD, which avoids the scaling problem that paper notes.

Second, we build the controlled testbed. CUB has attribute and part-keypoint annotations, so concept regions are defined. We add realistic correlated noise in two ways. In the first, we transplant real SIDD noise residuals (noisy minus ground truth), matched by local brightness to approximate their signal dependence. In the second, we simulate raw Poisson-Gaussian noise, then demosaicing and JPEG compression, the processing chain that correlates noise in Guo et al.. Having both matters because neither is perfect: the transplanted residuals are real but not exactly matched to CUB's content, and the simulation is matched but idealized.

Third, we move to microscopy. On CHAMMI, the benchmark itself flags models learning technical artifacts instead of morphology. We use label-free, ICE-style non-negative concept maps, since CLIP vocabularies are unreliable for cell images. We score accuracy with CHAMMI's 1-nearest-neighbour protocol. The fluorescence noise statistics that set the stride come from FMD.

Baselines

The naive reference is a SALF-CBM-style spatial concept model trained on clean CUB and tested on the noisy versions. It shows how much the problem matters at all. We know of no reported number for any concept bottleneck on noisy CUB or noisy microscopy, so we will measure this ourselves.

The strongest competitor is denoise-then-explain. AP-BSN with R³ refinement runs on sRGB inputs; it reports 35.97 dB PSNR on SIDD against 34.78 dB for R2R. FM2S runs on microscopy, where it reports a 1.4 dB margin over CVF-SID. The same spatial concept model is then trained on the denoised outputs. This baseline is dangerous because it removes most noise energy before the concept layer ever sees it. The denoisers use their released training recipes on the same noisy training images, and the concept model gets the same backbone, epochs and hyperparameter search as ours.

The second competitor is training with realistic correlated-noise augmentation, drawing from the same noise sources our method uses. This isolates whether cross-view agreement adds anything beyond exposure to noise.

We also compare against AIM. Its learned masks suppress spurious features, and it reports roughly 30% gains in energy pointing game on Waterbirds. We retrain it on the noisy data with its reported settings.

Two ablations isolate the key ingredient. The first replaces the pixel-shuffle views with two augmentations of the same pixels, so the noise stays shared between views. If this matches the full method, noise independence is not what matters. The second removes non-negativity.

Beating the field means the following. At clean CUB accuracy within one point of the naive model, and noisy-test accuracy at least equal to the augmentation baseline, we aim to reduce noise-realization concept variance and noise leakage below both denoise-then-explain and augmentation. We report the full accuracy-versus-stability curve across strides, not a single operating point.

Why not just…?

Why not just denoise the image first? This is the real threat, and the proposal is set up to lose to it honestly if it should. There are two reasons to expect it falls short. AP-BSN itself documents residual correlated structure after denoising and relies on multi-pass R³ refinement to suppress it. And any residual artifact the denoiser produces becomes part of the concept model's input with no mechanism stopping the concept layer from using it. The explanation then describes the denoiser's output rather than the measurement. The noise-realization variance experiment on FMD and SIDD repeated captures settles this directly. If denoise-then-explain matches our variance and leakage numbers, the paper's contribution becomes the measurement protocol plus that negative result about the method.

Why not just train with realistic noise augmentation? Augmentation teaches the classifier's output to tolerate noise on average. It does not constrain where a concept map draws its evidence, so a map can still flicker on noise while the final label stays correct. The shared-noise ablation and the augmentation baseline together separate "saw noise during training" from "was forced to agree across independent noise." We expect the gap to show in leakage and variance rather than in accuracy.

Why not just shrink or prune the concept predictor? The locality paper already shows pruning reduces leakage at a severe accuracy cost. We include a width sweep of the naive model so its accuracy-versus-leakage curve can be compared directly with ours.

Why not just enforce non-negativity? The physical-recovery literature notes that non-negativity helps against independent noise but is weaker when noise is correlated, because smooth positive noise blobs resemble genuine positive sources. We expect non-negativity alone to improve readability but not leakage, and the no-consistency ablation tests exactly this.

Decisive checks

The first check is whether noise-realization concept variance is substantially above zero for the naive and augmented models on real FMD and SIDD repeated captures. If it is near zero, the problem is not real and the project stops there.

The second check is whether our variance and leakage fall below denoise-then-explain at matched accuracy, with the gain disappearing in the shared-noise ablation.

The third check is whether the whitened stability score ranks models the same way as the empirical variance, while the unwhitened version does not. That would show white-noise stability analysis misjudges explanations on real sensors.

Secondary checks are necessity and sufficiency of concept maps under masking, and CHAMMI out-of-distribution 1-nearest-neighbour score. The reported reference points there are 0.537 for the HyperNet scratch model and 0.500 for FixedChannels; we would count as successful not dropping below them while stabilising the maps.