Relevance Propagation as Entropic Optimal Transport on Filippov Neural Networks: Sinkhorn Propagation and Region Relevance in Image Classifiers

Olivier Gibaru, Eric NyiriarXiv 2026 · arXiv

TL;DR

This paper provides a rigorous mathematical foundation for Layer-wise Relevance Propagation (LRP), proving it is not just a heuristic but the solution to a variational problem rooted in entropic optimal transport. By formalizing neural networks as graphs of Filippov operators, the authors extend attribution methods to modern, discontinuous architectures like hard-routed Mixture-of-Experts (MoE) and introduce a new, highly faithful rule called Sinkhorn Relevance Propagation.

What Problem Does This Paper Address?

Explainable AI (XAI) aims to peek inside the "black box" of neural networks to understand why they make specific predictions. Two popular approaches are gradient-based methods (sensitivity) and relevance propagation (attribution). While relevance propagation (specifically LRP) is widely used, it has long been criticized as a heuristic based on local conservation laws rather than a formal optimization principle.

Furthermore, standard XAI tools struggle with modern architectures that aren't smooth. For example, hard-routed Mixture-of-Experts (MoE) models—increasingly common in large language models—feature discrete switching logic that creates discontinuities. At these "routing boundaries," classical gradients are undefined, making traditional attribution methods unreliable or mathematically unsound.

Methods

The authors bridge these gaps using several novel techniques:

  1. Filippov Neural Networks: They redefine feed-forward networks as directed acyclic graphs (DAGs) of Filippov operators. This measure-theoretic approach allows for a generalized-gradient calculus that remains well-defined even when the network map jumps or is non-differentiable.
  2. Entropic Optimal Transport (EOT): The core contribution is showing that LRP is exactly the entropic projection of a "backward transition law" onto the set of divergence-free flows. In this view, different LRP rules ($LRP_0, LRP_{z+}, LRP_{\alpha,\beta}$) simply correspond to different choices of a "reference measure."
  3. Sinkhorn Relevance Propagation (SRP): By applying entropy edge-by-edge rather than across transition laws, the authors derive a new attribution rule computed via Sinkhorn-type iterations.
  4. Schrödinger Bridges: To explain specific regions of an image (e.g., "why did the model look at the dog's head?"), they use the framework of Schrödinger bridges to condition the relevance flow on specific input subsets.
Figure 1: The hard-routing MoE example: Φ jumps from −1 to 1 across the routing boundary z = 0, where g1(z) = g2(z). The Filippov regularization F[Φ](0) fills this jump with the whole segment [−1, 1] (dotted, in violet) rather than either endpoint alone, and the Filippov extension F[Φ](0) is the segment joining the pairs (Φ(k)(0), DΦ(k)(0)) of the two branches, (1, 1) and (−1, 2).

Datasets

The authors evaluated their framework on standard image classification benchmarks:

  • ImageNet: Used to test VGG-16 and ConvNeXt-Tiny architectures.
  • Imagenette: A smaller 10-class subset of ImageNet used for detailed faithfulness metrics (ABPC).
  • Oxford-IIIT Pet: Used for region-based relevance evaluation, utilizing the dataset's animal "trimaps" and head bounding boxes.

Key Insights and Ideas

  • LRP is Variational: The "aha" moment is that LRP isn't just a redistribution trick; it’s finding the "most likely" flow of information that satisfies conservation constraints, where "most likely" is defined by minimizing relative entropy to the forward pass.
  • Backprop as Optimal Control: The authors show that standard backpropagation can also be characterized variationally as the adjoint equation of a Lagrangian relaxation in a discrete-time optimal control problem.
  • Hebbian Relevance: They introduce a new rule, $LRPN_{\alpha,\beta}$, where weights vary based on node activity. This follows a Hebbian-inspired principle: more active nodes send more relevance to their excitatory contributions.
  • Unified DAG Framework: Their formalization covers everything from simple ReLUs to complex Transformer blocks with residual connections and hard-routing (see Figure 2).
Figure 2: A causal Transformer block on three tokens, drawn as a directed acyclic graph with input nodes it (the tokens) and output nodes ot. Each node stands for a group of neurons (one d-dimensional vector per token); expanding it coordinate by coordinate gives a DAG in the sense of Definition 2.3. The attention node at receives the query qt and the keys and values ks, vs of every earlier token s ≤t (solid orange edges), so the in-degree grows with the context length; dashed edges are residual connections skipping one or several stages, which prevent the graph from being layered (Definition 2.4). The double-circled node ft is the feed-forward sub-layer, which becomes a switching operator in the sense of Definition 2.1 when it is a hard-routed Mixture- of-Experts. CNN, ResNet, DenseNet, Inception and U-Net architectures give DAGs of the same kind.

Key Results

  • Mathematical Precision: On VGG-16, the proposed rules satisfy conservation identities to single-precision round-off error.
  • Robustness: $LRP_0$ (Gradient $\times$ Input) was found to be highly sensitive to noise, with a tiny perturbation ($10^{-5}$ std dev) shifting the map by 1.5%. In contrast, the EOT-based rules were 15x-25x more stable.
  • Superior Faithfulness: On the modern ConvNeXt-Tiny architecture, the Hebbian-weighted $LRPN_{\alpha,\beta}$ rule achieved the highest faithfulness (measured by ABPC) among all tested rules.
  • Region Specificity: Using Schrödinger bridges, the authors showed that in VGG-16, the animal's head receives roughly 2.5 times more relevance per pixel than the body, and that these two regions are processed through largely distinct hidden channels.

Limitations and Future Work

  • Computational Overhead: Sinkhorn Relevance Propagation (SRP) requires iterative sweeps, making it significantly slower (22s per image on an NVIDIA T1000) compared to standard LRP (0.16s).
  • Scaling to LLMs: While the theory is designed to handle the discontinuities found in MoE-based Large Language Models, the empirical evaluation in this paper is focused on image classifiers. Applying these tools to token-based attribution in LLMs is a clear next step.
  • Soft Constraints: The authors suggest that future work could relax the strict conservation constraint into a "soft" penalty using unbalanced optimal transport, which might be more appropriate for certain modeling goals.

Summary updated

Abstract

Relevance propagation, which redistributes a network's prediction backward through the layers, is usually treated as a heuristic. We show that it is the solution of a single variational problem: Layer-wise Relevance Propagation (LRP) is the divergence-free flow on the active computational graph of zero relative entropy to the backward transition law of the forward activation flow; for a non-negative reference, this is a relative entropy between Markov measures on backward paths. The rules LRP$_0$, LRP$_{z^+}$ and LRP$_{α,β}$ follow from this principle by the choice of a reference, and are instances of one rule, LRPN$_{α,β}$, whose weights may vary from node to node, for instance with the node's activity, in the spirit of Hebb's principle. The principle rests on a measure-theoretic formalization of feed-forward networks as directed acyclic graphs of Filippov operators, which covers piecewise-smooth and discontinuous architectures (e.g. hard-routed Mixture-of-Experts) with a Clarke-Filippov generalized-gradient calculus; it also characterizes backpropagation as the adjoint equation of a Lagrangian relaxation. On networks of ReLU and max-pooling nodes, LRP$_0$ is Gradient$\times$Input, with an explicit jump across a kink where LRP$_{z^+}$ stays continuous. Taking the entropy edge by edge yields a new rule, Sinkhorn Relevance Propagation. Aggregated over an input region, the flow of a non-negative reference tells, with exact accounting, which hidden units carry its evidence, the conditioned flow being a Schrödinger bridge. On VGG-16, the conservation identities hold to round-off, and the rules are told apart by the sign of their maps, the share of the output left to the bias nodes, faithfulness, localization and class dependence; on ConvNeXt-Tiny, LRPN$_{α,β}$ conserves the output and its Hebbian weights give the highest ABPC of the rules tested.
BibTeX
@misc{gibaru2026relevance,
  author = {Olivier Gibaru and Eric Nyiri},
  title = {{Relevance Propagation as Entropic Optimal Transport on Filippov Neural Networks: Sinkhorn Propagation and Region Relevance in Image Classifiers}},
  year = {2026},
  eprint = {2611.05594},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05594},
}
Figure