Explaining Instability Predictors Without Fake Silence

Deep models that read raw time series, such as audio classifiers and tokamak instability forecasters, are usually explained by deleting a piece of the input and watching the prediction change. For non-stationary signals this deletion is itself a corruption. Zeroing a stretch of audio, or of a magnetic-probe trace, creates an abrupt edge and an unnatural silence. The model may react to that artifact rather than to the missing content, and the resulting explanation then depends on which filler value was chosen. We propose replacing "delete" with "replace with what the surrounding signal predicts." A time-varying autoregressive model (a linear predictor whose coefficients drift smoothly over time) is fit to the context on both sides of the removed segment. Plausible continuations are drawn from it to fill the gap, and the attribution of the segment is how much the model's output depends on what the segment contained beyond what its neighbors already implied.

The same tool serves two settings that share a single failure. In audio, the masking problem has been demonstrated directly. In fusion, opaque predictors of plasma instabilities are a stated safety concern. Tokamak operating regimes also supply a natural test that audio lacks: a known physical shortcut, the periodic "sawtooth" core crashes, can be switched off by holding the central safety factor near one. We use that switch to check whether an explanation correctly predicts when a model will fail.

Core Claim

Reliable Local Explanations for Machine Listening showed that occluding spectrogram patches injects sharp spectral discontinuities that networks respond to. It also showed that saliency maps change drastically between zero-masking and mean-value masking. Our claim is that this baseline sensitivity comes from forcing the perturbed input off the signal's local statistics. A replacement drawn from a context-conditioned time-varying autoregressive (TVAR) model removes most of the sensitivity, because it keeps the edges continuous and the local spectral envelope intact. It also gives the attribution a precise meaning: the information in a segment that is not predictable from its surroundings. We define the attribution of segment $S$ of input $x$ as

$$\phi(S) = f(x) - \mathbb{E}_{\tilde{x}_S \sim p_\theta(\cdot \mid x_{\setminus S})}\left[ f(x_{\setminus S}, \tilde{x}_S) \right],$$

where $f$ is the model's output score, $x_{\setminus S}$ is the signal outside $S$, and $p_\theta$ is a TVAR model fit only on $x_{\setminus S}$. The TVAR coefficients are expanded in a small set of smooth temporal basis functions, following the time-varying linear prediction of Time-varying pre-emphasis and inverse filtering of speech and the multiscale TVAR tests of Time-varying autoregressive tests for multiscale speech analysis. Both papers supply the method; neither uses it for attribution.

The model order and basis smoothness determine what counts as "removed." A low-order, smooth TVAR keeps only the coarse spectral envelope and removes transients. A higher order keeps more. We treat this as an explicit, reportable knob rather than hiding it inside a filler constant.

Research Program

The work has three parts. The first is the replacement sampler. It uses two-sided TVAR gap interpolation on waveforms, with a forward and a backward predictor blended across the gap and residuals drawn from the context's estimated excitation, and it returns $K$ draws per segment. Since the perturbation happens in the time domain and features are recomputed afterwards, spectrogram-input models never see a mask-and-invert artifact. This sidesteps the phase-inconsistency problem noted for frequency-domain explanations.

The second part is a ground-truth audio testbed. Classifiers are trained with a deliberately planted shortcut: a class-correlated short broadband burst, resembling the plosive transients that Raw Speech-to-Articulatory Inversion by Temporal Filtering and Decimation notes stationary extractors discard. The burst sits at a known location. A raw-waveform model (Sample-level Deep Convolutional Neural Networks for Music Auto-tagging Using Raw Waveforms) and a spectrogram model are trained on environmental-sound tasks managed through DCASE-Models. For an emotion-recognition setup, we follow the attentive-pooling pipeline of Explainable Speech Emotion Recognition Through Attentive Pooling. Because the shortcut's location is known, an explanation's correctness can be scored, which addresses the missing-ground-truth problem.

The third part is a plasma transfer study, which is the stretch goal. Interpreting AI for Fusion applied Shapley attribution to Deep Survival Machines that forecast $n=1$ tearing-mode onset from DIII-D plasma profiles. Tearing modes are the instabilities in which magnetic surfaces tear into islands that leak heat. We build the analogous predictor on fast fluctuation signals: magnetic-probe and electron-cyclotron-emission traces, which are as non-stationary as speech. We explain it with the proposed method.

Axial safety-factor pinning enters as the evaluation regime. The Formation and Stationary Maintenance of Peaked Pressure Profiles in Turbulent Tokamak Plasmas showed in simulation that holding the on-axis safety factor $q_0 \approx 1$ suppresses sawtooth crashes. The safety factor counts toroidal turns of the field per poloidal turn. That work did not study transients and assumed a stationary density profile. Sawtooth crashes are the dominant transient in core temperature signals, so a predictor trained mostly on sawtoothing discharges may lean on them.

Sawteeth can physically seed tearing modes, so such reliance is not automatically spurious. That is exactly why the regime is useful. If the explanation says the model depends on crashes, then performance should degrade on sawtooth-free discharges in proportion to that dependence, and this prediction can be falsified either way. Access to tokamak data at this resolution is the main risk. If it fails, the audio testbed stands alone and the plasma study shrinks to a synthetic check. In that fallback, sawtooth-like relaxation oscillations are injected into or removed from recorded fluctuation traces.

Baselines

The naive reference is occlusion with zeros, and separately with mean training values, as used in spectrogram LIME/SHAP-style pipelines such as Iterative Feature Boosting for Explainable Speech Emotion Recognition. These run with the same segment grid, model, and number of forward passes as our method. The occlusion baselines are the setting where Reliable Local Explanations for Machine Listening documented instability. Beating them means higher attribution mass on the planted shortcut and substantially higher agreement between attributions computed under different filler settings.

The strongest existing competitor is the reliability-oriented local explanation procedure from that same machine-listening paper. We run it with its own recommended settings and an equal forward-pass budget. It is the method that most directly targets our failure mode. I found no reported numerical result on Lacuna for planted-shortcut localization or cross-baseline attribution agreement for any method, so all baseline numbers will be produced here under one protocol. For the plasma side, the file reports no quantitative metric for the Shapley analysis in Interpreting AI for Fusion. We will reimplement Shapley with its background-sample baseline on our fluctuation predictor.

A third baseline is a learned neural inpainter with the same context window, trained per dataset, which fills gaps generatively. The key ablation isolates the time-varying ingredient: a single stationary autoregressive model fit to the whole context, which matches our method in everything except drift. A second ablation fits the TVAR model including the segment itself. This should leak the content and shrink attributions toward zero, which confirms that fitting on the context alone matters.

Why not just…?

Why not just use gradient methods like integrated gradients and avoid masking altogether? Integrated gradients still requires a reference input, and a silent or zero reference reintroduces the same off-distribution problem along the path. Gradients of raw-waveform models are also notoriously noisy at the sample level. We include it as a baseline. If it matches our localization and baseline-agreement scores, the sampler is unnecessary and we will say so.

Why not just use an intrinsically interpretable model? Embedded factorization (Tackling Interpretability in Audio Classification Networks With Non-negative Matrix Factorization), prototype matching (AudioProtoPNet), and listenable time-domain separation (LMAC-TD) avoid post-hoc perturbation. However, they change or add a model, and constrained architectures can lose accuracy. Deployed fusion predictors and spoofing detectors (Uncovering the Deceptions) are often fixed black boxes. Our method audits the model that is actually running. We do not claim it beats a model that is interpretable by construction.

Why not just fill gaps with a large neural inpainter? This is the real threat. A strong generative infill could match our scores. Its failure modes are different: it can hallucinate class-typical content, which would make attributions reflect the inpainter's priors, and it needs training data from each domain, which plasma diagnostics rarely provide. The inpainter baseline settles the question. If it matches TVAR on audio, our remaining argument is cost and a transparent statistical definition of "removed," and the plasma regime, where data are scarce, becomes the decisive comparison.

Why not just use a stationary AR or noise filler? If the drift in TVAR does not matter, the stationary ablation will show it. We expect it to matter specifically for segments spanning transitions such as onsets, formant glides, and crash ramps, and we will report results stratified by local non-stationarity measured with the TVAR tests of Rudoy et al..

Proposed Experiments

On the audio testbed, the primary metrics are three. The first is the fraction of positive attribution mass that falls inside the planted burst. The second is the rank correlation between attribution maps computed under different filler settings: zero versus mean for occlusion, and different seeds and model orders for TVAR. The third is the "artifact response": the prediction change when a segment containing no class information is perturbed, which should be near zero for a faithful method. We vary shortcut strength and segment length and report results separately for raw-waveform and spectrogram models.

The decisive audio check is whether TVAR replacement raises cross-setting agreement substantially above occlusion while also localizing the shortcut. If agreement improves but localization does not, the method is merely smoothing.

On plasma data, the decisive check is predictive. We compute the share of each tearing-mode forecast's attribution that falls on sawtooth crashes in training-regime discharges. We then measure the drop in the predictor's discrimination when it is moved to sawtooth-free discharges, such as those with $q_0$ held near or above one. A useful explanation method should rank models or training variants by that drop. We train variants with sawtooth segments filtered out to create spread in reliance. We compare how well each attribution method's sawtooth share predicts the measured drop. If occlusion-based Shapley predicts the drop as well as ours, the stationarity argument does not carry over to fusion signals, and that negative result is itself worth reporting.