Do Neural-Field Explainers Remove Hyperparameter Sensitivity or Just Move It?

Saliency methods highlight the pixels a classifier relied on. They are known to be fragile: changing the random seed, the mask resolution, the blur radius or the number of optimization steps can change the heatmap as much as changing the model would. A recent line of work parameterizes the explanation mask as an implicit neural representation, meaning a small network that takes a pixel coordinate and returns a mask value. These masks are continuous and can be conditioned on the requested mask area. The published method is itself a trained network, with its own learning rate, depth, coordinate encoding and seed, and nobody has measured whether it is more stable than the grid-based optimizer it replaces.

We propose to measure this directly. We also test one specific mechanism by which a neural-field explainer could be more stable. If an explainer hyperparameter (mask area, and a new axis for mask smoothness scale) is turned into an input of the mask network, one set of shared weights must produce a coherent family of masks. That sharing may suppress configuration noise that separate grid optimizations cannot suppress. The deliverable is a sensitivity protocol that separates "the explanation changed because the model changed" from "the explanation changed because the explainer's settings changed." Applying it gives an answer that is useful whether it comes out positive or negative.

Core Claim

The quantity that matters is a ratio. The numerator is how strongly an explanation responds to changes in the model. The denominator is how strongly it responds to changes in the explainer's own configuration. We call this the signal-to-configuration ratio. It follows the causal framing of On the Relationship Between Explanation and Prediction: A Causal View, which treats explainer settings as outside interventions that can decouple the explanation from the prediction.

For an image $x$, let $D_{\text{model}}(x)$ be the mean dissimilarity between maps produced for the trained model and for weight-randomized versions of it, with the explainer configuration held fixed. Let $D_{\text{config}}(x)$ be the mean dissimilarity between maps for the trained model across sampled explainer configurations. The ratio is

$$R(x) = \frac{D_{\text{model}}(x)}{D_{\text{config}}(x)}.$$

Dissimilarity is one minus SSIM, with one minus Spearman rank correlation as a second measure. An explainer can make its maps stable by making them uninformative; a constant map has zero configuration spread. The ratio penalizes that, which a plain stability score does not. The idea mirrors how The Meta-Evaluation Problem in Explainable AI scores evaluation metrics by noise resilience and reactivity to model corruption. We apply that logic to explainers rather than to metrics.

Our hypothesis is specific. Among mask-optimization explainers, one that takes area and smoothness as inputs to a single coordinate network will have a higher $R$ than grid masks optimized separately per setting, at matched compute. The gain should come from the shared conditioning, not from continuity alone.

Research Program

The starting point is Generating Visual Explanations from Deep Networks using Implicit Neural Representations. That paper replaces the grid mask of extremal perturbation (find the smallest blurred-out region whose preservation keeps the class score high) with a network $m_\theta(x, y, a)$. It reports that the resulting masks track the requested area $a$ monotonically, which grid optimization does not. It also acknowledges that it explains one black box with another, and it runs no sensitivity analysis. That paper supplies our method. Its open limitation defines our question.

We add one conditioning variable $s$, the spatial scale of the mask. We borrow the device from Learning Continuous Image Representation With Local Implicit Image Function, where a "cell size" input tells a coordinate decoder how large an area each query covers, so that one model serves every output resolution. In our setting, $s$ replaces the hand-picked mask grid resolution or blur radius. Brunke et al. show this choice largely determines which features a perturbation method can see, in Evaluating Input Perturbation Methods for Interpreting CNNs and Saliency Map Comparison. The explainer then returns a whole surface of masks indexed by $(a, s)$. That surface can be reported directly, or averaged over a declared range of $(a, s)$ so that no single setting is chosen after the fact.

The protocol treats every remaining setting as a nuisance and samples it. For the neural-field explainers these are seed, learning rate, iteration count, network width and depth, and coordinate-encoding frequency. For the grid explainers the analogues are seed, learning rate, iterations and grid resolution. The fill used to "remove" pixels (blur, gray or black) is a known confound, since Brunke et al. show there is no neutral baseline. We cross it with every method and report it as a separate factor rather than claim to solve it.

We decompose the variance of each map into these factors with a factorial analysis. This shows where the instability lives for each explainer: in the optimizer, in the architecture, or in the fill. It is the direct test of "removed versus displaced."

Baselines

The naive reference is standard grid-based extremal perturbation run at its default settings and then swept over the same nuisance factors. This is the method the neural-field approach claims to improve on. On PASCAL VOC 2007, the neural-field paper reports mean precision 0.44 for its method against 0.46 for grid extremal perturbation, a difference it describes as not statistically significant (source). Localization alone therefore does not separate the two.

The strongest existing method is the area-conditioned neural-field explainer exactly as published. It reports mean precision 0.68 on ImageNet-S50 (752 images, 50 classes), above RISE and Grad-CAM (source). For its stability, no reported result was found on Lacuna. We reimplement it and must reproduce roughly that precision before any stability comparison counts.

The most dangerous competitor is a compute-matched ensemble of grid extremal perturbations. It averages masks over seeds and over several grid resolutions, using the same GPU-minutes per image as one neural-field fit (about three minutes per image, as reported). Averaging over configurations lowers configuration variance by construction, so this is the baseline that could make the neural-field explainer unnecessary.

Two ablations isolate our key ingredient. The first uses separate unconditioned coordinate networks, one per $(a, s)$ setting. This keeps continuity but removes weight sharing across settings. The second is a conditioned network with very low encoding frequency. It behaves almost like a smooth coarse grid and isolates what continuity itself contributes.

As external reference points we include RISE, Grad-CAM and SmoothGrad. We also include an adversarially trained ResNet as the target model. SAM: The Sensitivity of Attribution Methods to Hyperparameters reports that explanations of robust models are 2.4x less variable than those of standard models, and that meaningful perturbation converges in about 10 steps on robust models versus remaining unstable after hundreds on standard ones. That gives a known ceiling for how much stability a model-side change buys.

All methods share the same images, the same target classifiers, the same removal fills, and the same tuning budget. Each method is tuned only on a held-out set of 100 ImageNet validation images, and each gets the same number of sampled configurations. Beating a baseline means a higher median $R$ over images, with a paired bootstrap confidence interval that excludes zero. Localization precision must stay within the baseline's interval, and faithfulness under insertion and deletion curves must not drop.

Why not just…?

Why not just average grid masks over seeds and resolutions? This is a real threat, and we treat it as the central comparison rather than a straw man. Ensembling reduces the denominator of $R$. It does not by itself preserve the numerator, and averaging disagreeing masks can wash out model-specific structure. The compute-matched ensemble baseline settles it. If the ensemble matches the conditioned network on $R$, localization and faithfulness, we report that weight sharing adds nothing beyond ensembling. That would be a useful negative result for a method the community has begun to adopt.

Why not just use SmoothGrad? SAM finds that SmoothGrad makes a standard model's gradients look like a robust model's while leaving sensitivity to its own noise level and sample count. SAM calls this a de-noising illusion. Our ratio is designed to expose exactly this, because smoothing that lowers configuration spread while also muting the response to weight randomization will not raise $R$. We include SmoothGrad so this is shown on our own measurements, not only asserted from SAM.

Why not just explain an adversarially robust model? SAM's 2.4x figure shows this works when the auditor controls training. Auditors of fixed third-party models cannot retrain, and the robust model is a different model from the one being audited. We keep the robust ResNet as a reference ceiling. If the conditioned explainer on a standard model closes a meaningful fraction of the standard-to-robust stability gap, that is the practical payoff.

Why not just standardize on a toolkit such as Quantus and take consensus? Quantus standardizes how metrics are run. MetaQuantus shows its faithfulness metrics disagree with one another and are themselves sensitive to their settings. So we use Quantus only to compute metrics. We first screen those metrics with MetaQuantus meta-consistency, and we report every faithfulness score across all three removal fills so that no single fill choice can be cherry-picked.

Proposed Experiments and Decisive Checks

The main study uses ImageNet-S50, because its pixel-level masks allow localization precision to be compared with the reported 0.68. PASCAL VOC 2007 serves as a replication. The target classifiers are a standard ResNet-50, its adversarially trained counterpart, and a VGG-16 to check architecture transfer.

A full sweep at three minutes per fit is costly. We therefore run a fractional factorial design of about 40 configurations per method on a 200-image stratified subset, and run default configurations on all 752 images. The subset sweep takes on the order of 400 GPU-hours per neural-field method and is feasible.

The numerator of $R$ uses cascading weight randomization of the top layers. As a second check that explanations separate genuinely different models, we use independently trained seeds of the same architecture, scored with the AXE metric from Evaluating the Ability of Explanations to Disambiguate Models in a Rashomon Set.

We also count how often each method fails to converge on a given image. We use the stability criterion of Brunke et al., who report that only 389 of 1,000 ImageNet validation images reach a stable RISE map under default settings.

Three outcomes decide the proposal.

The first is whether the conditioned network's median $R$ exceeds the compute-matched grid ensemble's at equal localization and faithfulness. If it does not, the stabilization claim fails.

The second is whether the unconditioned per-setting coordinate networks fall back to the grid baseline's $R$. If they do, weight sharing across settings, not continuity, is the active ingredient. If they do not, continuity is.

The third is the variance decomposition, which shows where each explainer's residual instability sits. If most of the neural-field explainer's spread moves into encoding frequency or learning rate, the method has displaced the sensitivity rather than removed it, and we will say so. Either way, the removal fill is reported as its own factor. We expect it to remain a large share of the variance for every method, which would mark it as the problem continuous masks do not touch.