Skip to content

Back to the catalog

Revisiting Explainable AI through Model-Independent Concept Dictionaries
arxiv.orgpaper

Revisiting Explainable AI through Model-Independent Concept Dictionaries

DictXAI is a new explainable AI method that defines interpretable concepts directly in the input domain using a dictionary of predefined elements, rather than relying on inherently-interpretable input features or architecture-specific internal abstractions as prior XAI methods do. It computes a sparse code of the model input and attributes predictions to dictionary elements, letting it work across learned image bases, analytically-defined ECG waveforms, and experimentally acquired dictionaries, and the authors show it can trace AI malfunctions like Clever Hans effects to specific data artifacts while aiding human-AI alignment on biomedical signals.

TL;DR A model-independent, dictionary-based XAI method (DictXAI) attributes predictions to interpretable input-domain concepts, working across image and ECG data to expose Clever Hans artifacts.
Open original ↗
Document

What it claims

DictXAI explains a trained model's predictions in terms of input-space dictionary atoms instead of pixels or hidden-layer concepts. Examples of atoms include Gabor filters at a given scale and angle, learned stroke primitives, amplitude-modulated sinusoidal (AMS) ECG waveforms, and measured microbial mass-spectrometry profiles.

Two design choices separate it from related work:

  • Dictionaries may be overcomplete and non-orthogonal. Fourier-, wavelet- and CartoonX-style methods are tied to fixed orthogonal transforms. An overcomplete dictionary can separate highly correlated variations, such as continuous scale shifts in a waveform.
  • No trained encoder. Unlike SAE or NMF concept methods, the explanation depends only on the original model and the explicit atoms.

The authors claim that removing the input-space atoms behind a Clever Hans strategy blocks that strategy "at its source." They say this stops the strategy from re-emerging through alternative internal circuits.

How it works

  1. Sparse code per sample. Solve the LASSO problem min_α ½‖Σ αⱼdⱼ − x‖² + λΣ|αⱼ| directly for each input, with no amortized encoder. The authors argue this avoids encoder capacity limits and out-of-distribution failures of the encoder. They note that computing α with an encoder instead would make the method essentially a sparse autoencoder.
  2. Lossless reconstruction. Add an instance-specific residual atom d₀ = x − x̂ with fixed coefficient α₀ = 1. The model then sees exactly x, so f(Σ_{j=0..K} αⱼdⱼ) = f(x). The prediction and the decision strategy are not distorted by reconstruction error.
  3. Attribute to atoms. Standard attribution methods carry over: - Occlusion: Rⱼ = f(x) − f(x − αⱼdⱼ). This extends to Shapley value sampling over atom coalitions. - Integrated Gradients: the path scales the coefficients as α(t) = tα. - LRP-0: pixel relevances are redistributed back to atoms as Rⱼ = Σᵢ ([αⱼdⱼ]ᵢ / xᵢ) Rᵢ, with the convention 0/0 = 0. - More stable LRP: merge the dictionary layer into the model's first linear layer so that LRP-γ can be used.

Because step 3 reuses existing attribution methods, the authors say DictXAI inherits their faithfulness properties, such as conservation and continuity.

Dictionary options and trade-offs:

TypeExample from the pageTrade-off stated
LearnedMNIST dictionary with K = 800 atoms, which come out as localized stroke contoursBest sparsity/reconstruction trade-off, but atoms have no parametric formula, which makes interpretation harder
AnalyticGabor filters (angle, scale); AMS waveforms for ECG QRS durationExplanations can be read along explicit axes
ExperimentalReference spectra of microbial isolatesPhysically grounded

Evidence

The fetched text does not include any quantitative results. The application studies are only named:

  • Clever Hans removal on images
  • Interpreting predicted QRS duration from ECG
  • Mass-spectrometry analysis

No metrics, baselines or dataset sizes are given beyond K = 800 for MNIST. The claim of being "more interpretable, actionable, and architecture-agnostic" than classical and concept-based XAI is not backed by numbers in this excerpt.

Limits

  • Page truncated. The text cuts off partway through Section 3.1, so the evaluation (Section 4) was not available. Read the full PDF before relying on the comparative claims.
  • Learned dictionaries need data. They assume plenty of samples; with scarce data the dictionary risks overfitting. Their atoms are also harder to interpret than parametric ones.
  • Per-sample cost. Each explanation requires solving a LASSO problem. The page gives no figures on runtime or on how to choose λ.
  • Residual atom d₀. It keeps the forward pass exact but can itself receive attribution. How much relevance lands on d₀ in practice is not reported in the excerpt.
  • Adoption details missing. No code, licence or install information appears on the page.

Organisations named here

  • Grégoire Montavon

Techniques named here

  • DictXAI

Related in research