DeepFaith: Unified Faithful Explainer

• The paper introduces DeepFaith, a unified framework that reframes explanation generation as a supervised learning problem with an optimal faithfulness objective.
• DeepFaith employs a multi-layer Transformer encoder to generate saliency maps in one forward pass, optimizing both pattern consistency and local correlation losses.
• Empirical evaluations across 12 tasks demonstrate that DeepFaith outperforms traditional methods with the best average rank on ten faithfulness metrics.

Searching arXiv for the cited DeepFaith papers to ground the article in current metadata. arxiv_search(query="(Guo et al., 5 Aug 2025) DeepFaith domain-free model-agnostic unified framework faithful explanations", max_results=5) arxiv_search(query="(Iwata et al., 2021) Training Deep Models to be Explained with Fewer Examples", max_results=5) Deep Architecture-Based Faith Explainer, abbreviated DeepFaith, denotes a domain-free, model-agnostic unified explanation framework under the lens of faithfulness that recasts explanation generation as a supervised learning problem grounded in a unified theory of faithfulness metrics. Rather than computing an explanation by repeatedly querying the target model at inference time, DeepFaith identifies a theoretically motivated optimal explanation objective, distills high-quality supervision from multiple existing explainers, and trains a neural explainer that outputs a saliency map in a single forward pass (Guo et al., 5 Aug 2025).

1. Definition and conceptual scope

DeepFaith is formulated around a familiar XAI difficulty: there is no ground truth explanation for a model decision, so explanation quality is typically assessed through surrogate criteria. The framework treats faithfulness as the central criterion and argues that the absence of a unified optimal explanation has prevented objective evaluation and optimization across methods (Guo et al., 5 Aug 2025).

The framework distinguishes two explanation families. A saliency explanation is written as

$S_f:\mathcal{X}\to[0,1]^n,$

producing a saliency vector $s=(s_1,\dots,s_n)$. A permutation explanation is written as

$\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$

producing an ordering of feature importance. The two are connected by

$$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$

and

$\Sigma(\pi)_{\pi(i)}=\frac{n-\pi(i)+1}{n}.$

This formalism is intended to cover image patches, tokens, or tabular features within one notation (Guo et al., 5 Aug 2025).

A central operational choice is the perturbation notation $x\setminus \mathcal{I}$, meaning that features in $\mathcal{I}\subseteq[n]$ are removed or replaced using noise, baseline substitution, or interpolation. Faithfulness is then expressed through how well an explanation aligns with the model’s perturbation response, using a perturbation effect $\Delta$, a preservation effect $\Delta^-$, and a correlation measure $\tau$ (Guo et al., 5 Aug 2025).

2. Unified faithfulness theory

DeepFaith’s technical core is a unified formulation spanning ten widely used faithfulness metrics. On the saliency side, the framework reformalizes Faithfulness Correlation (FC), Faithfulness Estimate (FE), Infidelity (INF), and Monotonicity Correlation (MC). Their shared principle is to correlate the saliency mass assigned to a perturbed subset with the model’s change under perturbation: $s=(s_1,\dots,s_n)$0 The paper emphasizes that these metrics differ in sampling scheme, but all encode the same principle: explanations should assign larger scores to feature subsets whose removal has larger impact on model output (Guo et al., 5 Aug 2025).

On the permutation side, the unified view covers Deletion Score (DEL), Insertion Score (INS), Negative Perturbation (NEG), Positive Perturbation (POS), Region Perturbation (RP), and Iterative Removal of Features (IROF). These metrics remove or insert features in ranked order and accumulate the effect. For example, the deletion score is written as

$s=(s_1,\dots,s_n)$1

The theoretical claim is that a single ordering principle underlies these ranking-based metrics as well (Guo et al., 5 Aug 2025).

The key theorem starts from an optimal saliency mapping

$s=(s_1,\dots,s_n)$2

The paper states that if such an $s=(s_1,\dots,s_n)$3 exists, then it is optimal under FC, FE, INF, and MC, and that the induced ranking

$s=(s_1,\dots,s_n)$4

is optimal for DEL, INS, NEG, POS, RP, and IROF as well (Guo et al., 5 Aug 2025). This makes the optimal saliency mapping a theoretical “ground truth” under the paper’s faithfulness lens.

3. Learned explainer architecture and supervision construction

DeepFaith implements the explainer as a deep model

$s=(s_1,\dots,s_n)$5

with a multi-layer Transformer encoder followed by a normalized linear layer that produces an $s=(s_1,\dots,s_n)$6-dimensional saliency vector (Guo et al., 5 Aug 2025). The explainer is therefore not tied to the target model’s internal mechanics at inference time; instead, model-specific faithfulness is absorbed during training.

A distinctive engineering contribution is the construction of supervised explanation signals from multiple existing explainers. For each sample, candidate explanations

$s=(s_1,\dots,s_n)$7

are first generated. The framework then applies two curation steps (Guo et al., 5 Aug 2025):

• Deduplicating: pairwise cosine similarities among candidate explanations are computed, duplicate groups are formed using a manually defined similarity threshold, and only the first explanation in each duplicate group is kept.
• Filtering: each retained explanation is evaluated by all ten faithfulness metrics; thresholds are computed metric-wise using a $s=(s_1,\dots,s_n)$8-quantile over deduplicated explanations, with direction reversed for metrics where lower is better; an explanation is kept only if it satisfies the threshold for every metric.

The resulting curated supervision set is

$s=(s_1,\dots,s_n)$9

This design is meant to avoid learning indiscriminately from weak or redundant explanation methods (Guo et al., 5 Aug 2025).

4. Optimization objective and inference behavior

DeepFaith is trained with two coupled losses. The Pattern Consistency loss aligns the explainer with the curated supervision: $\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$0 The Local Correlation loss directly optimizes the unified faithfulness criterion: $\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$1 where

$\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$2

The combined objective is

$\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$3

The training schedule starts with $\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$4, so optimization is initially dominated by pattern consistency, then shifts toward local correlation as $\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$5 saturates (Guo et al., 5 Aug 2025).

A defining property is inference decoupling. After training, DeepFaith takes only the input $\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$6 and produces

$\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$7

in a single forward pass. It does not need access to $\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$8 during inference (Guo et al., 5 Aug 2025). This is a substantive departure from post-hoc explainers that must repeatedly query or backpropagate through the target model at test time. A plausible implication is that DeepFaith is designed to trade substantial offline training cost for fast deployment-time explanation generation.

5. Empirical evaluation and reported results

The reported evaluation spans 12 explanation tasks across 6 datasets and 6 model families (Guo et al., 5 Aug 2025). The datasets and model families are:

• Image: ImageNet and UCSD OCT Retina, each with ResNet50, EfficientNet-b0, and DeiT
• Text: IMDb and AGNews, each with LSTM and Transformer
• Tabular: NHANES Age Prediction and Wholesale Customers Data, each with MLP

The supervision pool is modality-specific. For image tasks it includes Occlusion, LIME, Kernel SHAP, DeepLIFT SHAP, Saliency, Input × Gradient, Guided Backprop, Grad-CAM, Score-CAM, Grad-CAM++, Integrated Gradients, Expected Gradients, DeepLIFT, and LRP. For text and tabular tasks it includes Integrated Gradients, Gradient SHAP, DeepLIFT, Saliency, Occlusion, Feature Ablation, LIME, and Kernel SHAP (Guo et al., 5 Aug 2025).

The main quantitative claim is that DeepFaith achieves the best average rank across all 10 metrics on all 12 tasks (Guo et al., 5 Aug 2025). Reported examples include:

Task	DeepFaith average rank	Baseline examples
OCT + DeiT	3.4	Integrated Gradients 7.8; Saliency 13.2
ImageNet + DeiT	4.4	Better than most baselines
IMDb + Transformer	2.7	Outperforms all compared methods
NAP + MLP	1.8	Best among all methods

The ablation study compares $\Pi_f:\mathcal{X}\to\mathfrak{S}_n,$9, $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$0 only, and $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$1 only. The combined objective consistently performs best; $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$2 alone is limited by the quality and diversity of baseline explanations, whereas $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$3 alone is difficult to optimize and may fail to converge effectively (Guo et al., 5 Aug 2025).

The paper also frames runtime as a two-phase trade-off. Training is expensive because explanation signal generation and explainer training are costly, but once trained, DeepFaith is significantly faster than sampling-based methods such as LIME, Kernel SHAP, and Occlusion, and is competitive with or faster than gradient-based methods in the reported setting (Guo et al., 5 Aug 2025).

6. Relation to adjacent “faithful-by-design” work and naming context

The name DeepFaith has an earlier usage in a different framework. In “Training Deep Models to be Explained with Fewer Examples”, DeepFaith refers to a method that jointly trains a prediction model $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$4 and an example-based explanation model $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$5 derived from Representer Point Selection, using a sparse regularizer with stochastic gates so that predictions can be faithfully explained by a small number of training examples (Iwata et al., 2021). That earlier DeepFaith is therefore an example-based co-training method rather than the later domain-free, model-agnostic explainer-learning framework.

The broader literature contains several methods that share a related concern with architectural or mechanistic faithfulness. FCVE explains a trained CNN by identifying minimum correct (MC) and minimum incorrect (MI) convolutional filters at the last convolutional layer and reconstructing the corresponding image with a decoder, so the counterfactual visual explanation is produced by manipulating the classifier’s internal representation rather than input pixels (Khan et al., 12 Jan 2025). FaCT combines B-cos layers with bias-free sparse autoencoders so that concepts are part of the forward pass, enabling exact decomposition of logits into concept contributions and tracing concept activations back to input pixels (Parchami-Araghi et al., 29 Oct 2025). ProtoPFaith replaces ProtoPNet’s heuristic prototype heatmaps with Shapley values over prototype similarity scores, aiming to satisfy axioms such as sensitivity, implementation invariance, completeness, dummy, linearity, and symmetry-preserving (Wolf et al., 2023).

These neighboring methods suggest two distinct but related traditions. One tradition, exemplified by FCVE, FaCT, and ProtoPFaith, seeks explanations that are faithful because they remain close to the model’s internal causal or computational structure. The 2025 DeepFaith framework instead learns a separate explainer from curated supervision and a unified faithfulness objective, then dispenses with access to the target model at inference time (Guo et al., 5 Aug 2025). This suggests a conceptual distinction between architecture-native faithful explanation and model-agnostic faithful explainer learning.

7. Limitations, practical constraints, and significance

Several limitations are explicitly stated or clearly implied by the framework description. DeepFaith depends on access to a sufficiently rich set of candidate explanation methods to generate supervised signals; its behavior is influenced by the filtering threshold $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$6, the deduplication similarity threshold, and the choice of similarity or correlation measure $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$7 (Guo et al., 5 Aug 2025). The learned output is a saliency vector, with permutation explanations derived from rankings of that vector rather than generated natively. Training cost is higher than for standard post-hoc explainers because supervision construction itself is expensive (Guo et al., 5 Aug 2025).

The paper also notes that the approach is only as good as the curated supervision. This suggests that if all candidate explanation methods are poor in a domain, signal quality may deteriorate even though the framework still optimizes the local correlation objective. A further plausible implication is that DeepFaith’s theoretical unification does not eliminate the practical importance of perturbation design, since the definition of $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$8, $$\mathfrak{P}(s)=\mathrm{argsort}_\downarrow\{s_1,\dots,s_n\},$$9, and $\Sigma(\pi)_{\pi(i)}=\frac{n-\pi(i)+1}{n}.$0 remains part of the faithfulness machinery.

Within the literature surveyed here, DeepFaith’s significance lies in shifting faithfulness from a purely evaluative criterion to a trainable target. Its central claim is not merely that explanations should correlate with perturbation effects, but that a single optimal saliency mapping can jointly satisfy a broad family of saliency-based and permutation-based faithfulness metrics, and that a neural explainer can be trained to approximate that mapping efficiently at deployment time (Guo et al., 5 Aug 2025).