Data-Driven Feature Interpretation Dictionary

• DDFID is a structured approach that organizes learned model features into interpretable dictionaries by enforcing sparsity and linking to semantic explanations.
• It leverages network activations, sparse dictionary learning, and post-hoc attribution methods to build compact bases that align functional performance with domain semantics.
• Its applications span mechanistic interpretability in deep classifiers, radiomics for breast cancer analysis, and diffusion-model analysis with both predictive and interventional capabilities.

Data-Driven Feature Interpretation Dictionary (DDFID) denotes a class of interpretability constructions in which features are organized into a dictionary whose entries are learned from data, selected from internal activations, or induced from post-hoc attribution patterns, and then linked to semantic explanations. In the cited literature, DDFID is not a single standardized formalism; rather, it appears as a recurring design pattern across joint interpretation architectures, sparse dictionary learning, mechanistic interpretability modules, diffusion-model analysis, and radiomics. The common structure is a compact or sparse feature basis together with a procedure that makes individual dictionary entries interpretable at local, global, or domain-semantic levels (Parekh et al., 2020, Braun et al., 2024, Wu et al., 2024, Shabalin et al., 30 May 2025, Gorji et al., 21 Jul 2025).

1. Conceptual scope and recurring structure

Across the relevant works, a DDFID-like construction has three recurrent components. First, it defines a feature inventory: hidden-layer attributes in FLINT, sparse codes in sparse autoencoders, dictionary features in DILA, activation atoms in diffusion-model decomposition, or radiomic features interpreted through SHAP. Second, it imposes a structural constraint such as sparsity, conciseness, nonnegativity, or dictionary overcompleteness, so that only a small subset of entries is active or salient for a given instance. Third, it supplies an interpretation channel: linear relevance decomposition, top-activating contexts, LLM summarization, activation maximization, or mapping to a clinical lexicon (Parekh et al., 2020, Wu et al., 2024, Shabalin et al., 30 May 2025, Gorji et al., 21 Jul 2025).

This shared pattern can be summarized compactly.

Work	Dictionary source	Interpretation mechanism
FLINT (Parekh et al., 2020)	Hidden-layer attribute activations $\Phi(x)$	Relevance scores, local/global interpretation, AM+PI
End-to-end sparse dictionary learning (Braun et al., 2024)	SAE dictionary $D$ and sparse code $z$	Top-activating tokens, ablation, resample patching
DILA (Wu et al., 2024)	Sparse token-level dictionary features	LLM odd-one-out test, short concept summaries, sparse code-to-label matrix
Diffusion-model dictionary learning (Shabalin et al., 30 May 2025)	SAE or ITDA atoms from residual-stream activations	Top-activating images, LLM auto-interpretation, steering
BM1.0 radiomics DDFID (Gorji et al., 21 Jul 2025)	Radiomic features not covered in CIFID	SHAP direction and magnitude mapped to BI-RADS–style descriptors

A plausible implication is that DDFID is best understood as an umbrella term for data-grounded feature lexicons rather than as a single algorithm. Some instances are intrinsically model-based and learned end-to-end; others are post-hoc semantic overlays on an already selected feature set.

2. Learned dictionary architectures in predictive models

In joint learning settings, the dictionary is embedded directly into the predictive architecture. FLINT starts from a deep classifier

$f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$

selects hidden layers indexed by $I=\{i_1,\dots,i_T\}$, concatenates their outputs into $f_I(x)\in\mathbb{R}^D$, and applies a shallow attribute-dictionary network $\Psi:\mathbb{R}^D\to\mathbb{R}^J$ to produce non-negative attribute activations

$$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$

A linear interpreter head then yields

$g(x)=h\circ \Psi \circ f_I(x),$

with $h(\Phi)=\mathrm{softmax}(W^\top\Phi)$. The total loss is

$D$0

where

$D$1

Here $D$2 enforces output-fidelity to $D$3, $D$4 is an input-fidelity autoencoder term through a decoder $D$5, and $D$6 combines diversity, conciseness, and an $D$7 control on raw activations. FLINT explicitly seeks “a small-size dictionary of high level attribute functions” and provides both local and global interpretation through relevance scores derived from the linear head (Parekh et al., 2020).

The relevance machinery is central to FLINT’s DDFID-like behavior. For a sample $D$8 and predicted class $D$9, the raw contribution of attribute $z$0 is $z$1, and the normalized relevance is

$z$2

Local interpretation is given by

$z$3

while global interpretation averages $z$4 over class-specific samples. Thus the dictionary is not merely a reconstruction basis; it is also a class-conditioned explanatory coordinate system (Parekh et al., 2020).

DILA instantiates a related but distinct architecture for high-dimensional multilabel prediction. Token embeddings $z$5 are passed through a dictionary learning autoencoder with

$z$6

The sparse feature matrix $z$7 is then projected to labels through a global dictionary-to-code matrix

$z$8

producing

$z$9

The combined objective is

$f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$0

with sparse autoencoder loss

$f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$1

In DILA, the dictionary is explicitly interpreted as a set of globally learned medical concepts, and the sparse matrix $f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$2 provides a global mechanistic relation between concepts and ICD codes (Wu et al., 2024).

3. Sparse dictionary learning and functional importance

A major DDFID line of work treats the dictionary as an overcomplete sparse basis for internal activations. In end-to-end sparse dictionary learning, a frozen network produces a layer activation $f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$3, and a sparse autoencoder defines

$f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$4

Unlike local SAEs trained only for reconstruction, the end-to-end objective minimizes KL divergence between the original model output distribution and the model with reconstructed activations inserted, together with an $f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$5 sparsity term and an optional downstream MSE term:

$f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$6

This formulation is designed so that the learned features are functionally important to the network rather than only descriptive of the dataset (Braun et al., 2024).

The quantitative comparison is central. On GPT2-small layer 6, with matched CE loss increase $f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$7, the reported runs are: local SAE with $f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$8, $f(x)=f_{l+1}\circ f_l\circ \cdots \circ f_1(x),$9, and 26,000 alive features; end-to-end SAE with $I=\{i_1,\dots,i_T\}$0, $I=\{i_1,\dots,i_T\}$1, and 22,000 alive features; and e2e+downstream with $I=\{i_1,\dots,i_T\}$2, $I=\{i_1,\dots,i_T\}$3, 15,000 alive features, and CE increase $I=\{i_1,\dots,i_T\}$4. Pareto curves further indicate that, for the same CE loss increase, local SAE uses more than twice as many features as e2e SAE, and for a given $I=\{i_1,\dots,i_T\}$5, local SAE has more than twice the CE increase. These results operationalize a DDFID not as a purely descriptive dictionary, but as a sparse basis constrained by model behavior (Braun et al., 2024).

A related sparse-dictionary formulation appears in large text-to-image diffusion models. There the objective is to reconstruct residual-stream activations $I=\{i_1,\dots,i_T\}$6 as

$I=\{i_1,\dots,i_T\}$7

using either a Top-$I=\{i_1,\dots,i_T\}$8 sparse autoencoder or Inference-Time Decomposition of Activations (ITDA). The SAE loss is written in general form as

$I=\{i_1,\dots,i_T\}$9

subject to $f_I(x)\in\mathbb{R}^D$0, while ITDA performs

$f_I(x)\in\mathbb{R}^D$1

by gradient pursuit and grows the dictionary by appending poorly reconstructed activations as new atoms. The reported implementation uses $f_I(x)\in\mathbb{R}^D$2, sparsity $f_I(x)\in\mathbb{R}^D$3, learning rate $f_I(x)\in\mathbb{R}^D$4, and 30,000 steps on 30 M tokens; PCA whitening is described as crucial to avoid dead latents because of heavy anisotropy in Flux 1 activations (Shabalin et al., 30 May 2025).

The diffusion-model results show that DDFID can support both interpretation and intervention. Fraction of Variance Unexplained is measured as

$f_I(x)\in\mathbb{R}^D$5

and for layer 18 both SAE and ITDA reach approximately $f_I(x)\in\mathbb{R}^D$6–$f_I(x)\in\mathbb{R}^D$7 FVU at 64 K atoms, with ITDA slightly outperforming SAE at modest sizes below 10 K atoms. For automated interpretability at layer 18 with $f_I(x)\in\mathbb{R}^D$8 K and $f_I(x)\in\mathbb{R}^D$9, mean judge accuracy is approximately $\Psi:\mathbb{R}^D\to\mathbb{R}^J$0 for SAE features, $\Psi:\mathbb{R}^D\to\mathbb{R}^J$1 for ITDA features, and $\Psi:\mathbb{R}^D\to\mathbb{R}^J$2 for individual MLP neurons. Steering is performed by modifying the first 5 of 7 diffusion steps via activation addition in a spatial window,

$\Psi:\mathbb{R}^D\to\mathbb{R}^J$3

This makes the dictionary a causal control surface as well as an explanatory one (Shabalin et al., 30 May 2025).

4. Post-hoc semantic dictionaries and domain lexicons

Not all DDFIDs are learned as latent bases. In BM1.0 for personalized breast cancer analysis, the DDFID is generated post hoc from model behavior. The task is binary classification of triple-negative breast cancer versus non-TNBC using late post-contrast DCE-MRI radiomic features from a multi-institutional cohort of 1,549 patients. The pipeline consists of N4 bias-field correction, resampling to uniform voxel spacing, z-score normalization, extraction of 108 standardized radiomic features via PyRadiomics, Variance Inflation Factor feature selection, Extra Trees classification, and SHapley Additive exPlanations for interpretation. The DDFID covers the 52 radiomic features not covered in the Clinically-Informed Feature Interpretation Dictionary, which already mapped 56 features to BI-RADS descriptors through literature and expert review (Gorji et al., 21 Jul 2025).

The mathematical basis is explicitly attributional. For a model $\Psi:\mathbb{R}^D\to\mathbb{R}^J$4 with $\Psi:\mathbb{R}^D\to\mathbb{R}^J$5 features, the SHAP value for feature $\Psi:\mathbb{R}^D\to\mathbb{R}^J$6 is

$\Psi:\mathbb{R}^D\to\mathbb{R}^J$7

and the local additive explanation model is

$\Psi:\mathbb{R}^D\to\mathbb{R}^J$8

VIF is computed as

$\Psi:\mathbb{R}^D\to\mathbb{R}^J$9

with features typically discarded when $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$0. In this setting, DDFID construction maps the sign and magnitude of SHAP values to BI-RADS–style semantic descriptors associated with TNBC or non-TNBC, such as round/oval shape, smooth margins, rim enhancement, irregular shape, spiculated margins, and heterogeneous enhancement (Gorji et al., 21 Jul 2025).

The resulting dictionary is explicitly semantic rather than latent. Representative entries include Major Axis Length favoring TNBC and being mapped to round/oval shape and smooth margins, Long Run Emphasis favoring non-TNBC and being mapped to heterogeneous enhancement, and Surface Area favoring non-TNBC and being mapped to irregular or complex surface. The best reported model combines VIF selection with Extra Trees Classifier and achieves an average cross-validation accuracy of $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$1; the abstract states this as $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$2, while the detailed description gives $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$3 across 5 CV folds (Gorji et al., 21 Jul 2025).

A related but distinct direction is Partial Information Decomposition of Features. PIDF introduces three metrics per feature: the mutual information shared with the target variable, the feature’s contribution to synergistic information, and the amount of this information that is redundant. Its stated aim is simultaneous data interpretability and feature selection, with case studies in genetics and neuroscience. However, the available description does not provide the explicit PID equations, an algorithmic estimator for the PID components, or a procedure for assembling those quantities into a DDFID. This suggests a conceptual proximity between dictionary-based interpretation and multi-component feature semantics, but the available text does not establish a formal DDFID construction for PIDF (Westphal et al., 2024).

5. Interpretation outputs, validation protocols, and reported performance

DDFID systems differ sharply in how they validate interpretability. FLINT emphasizes fidelity, conciseness, and visualization. Its interpreter fidelity is the fraction of inputs for which $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$4 and $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$5 agree; on MNIST, Fashion-MNIST, CIFAR-10, and QuickDraw, the interpreter attains more than 90 percent fidelity to the predictive model. Conciseness is measured as the average number of attributes per sample above the threshold $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$6, and FLINT-g produces 2–4 attributes per sample versus 6–10 for SENN. A shuffling test shows that randomly permuting $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$7 drops FLINT-g accuracy by approximately 75 percentage points but barely affects SENN, and a QuickDraw user study reports that 77.5 percent of correct textual descriptions strongly agreed with AM+PI visualizations, while maliciously incorrect descriptions strongly disagreed 83.7 percent of the time. Training with the interpretability loss is reported not to harm the base model’s accuracy, within $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$8 percentage points (Parekh et al., 2020).

End-to-end sparse dictionary learning validates interpretation through mechanistic faithfulness to model behavior. The central metrics are CE loss increase after SAE insertion, $$\Phi(x)=[\phi_1(x),\dots,\phi_J(x)]^\top=\Psi(f_I(x)).$$9 as the average number of nonzero features per token, and the number of alive dictionary elements that ever activate on 500 k tokens. The method is presented as a Pareto improvement over standard SAEs: it explains more network performance, requires fewer total features, and requires fewer simultaneously active features per datapoint, with no cost to interpretability. The accompanying geometric analysis measures maximum cosine similarity between atoms, cross-seed stability, cross-type similarity, and UMAP cluster structure, with the reported interpretation that local SAEs contain more near-duplicate “split” features and more functionally irrelevant distinctions (Braun et al., 2024).

DILA uses both human and automated concept validation. Human evaluations indicate that its sparse embeddings are more human understandable than dense counterparts by at least 50 percent. Automated dictionary feature identification collects top-activating tokens for each feature, subjects them to a 5-way odd-one-out test using a medical LLM, and then summarizes coherent features into labels of at most 8 words. The DDFID row may include the feature index, concept label, top-$g(x)=h\circ \Psi \circ f_I(x),$0 activating tokens, and selected strong code activations from the sparse global matrix. This yields both global interpretability, through the code-to-feature matrix, and local interpretability, through the activation of specific tokens in specific notes (Wu et al., 2024).

In diffusion-model dictionary learning, evaluation combines reconstruction, automated interpretation, and intervention. Interpretability is measured by an LLM “judge” that decides whether a candidate explanation matches highlighted held-out images, and the average 0–1 judgments define an autointerp accuracy. Qualitatively, SAE atoms tend to capture concrete objects such as “a baseball cap” or “a hand holding a pencil,” whereas ITDA atoms are often more abstract, such as “a textured background” or “a blob of warm color.” Steering experiments further validate that certain atoms encode manipulable concepts, although steering artifacts can appear when the prompt and layer state are too orthogonal to the feature being added or subtracted (Shabalin et al., 30 May 2025).

6. Conceptual status, limitations, and open issues

The cited literature supports a broad but non-uniform understanding of DDFID. In some works, the dictionary is a learned latent basis tied directly to internal activations; in others, it is a semantic overlay on selected radiomic features; in still others, it is a sparse concept layer connecting tokens to labels. This suggests that DDFID currently functions as an umbrella label for data-grounded feature lexicons rather than a universally agreed technical specification (Wu et al., 2024, Gorji et al., 21 Jul 2025).

The main methodological tension concerns what kind of “interpretation” the dictionary is supposed to guarantee. Reconstruction-focused dictionaries can recover structure in activations, but the end-to-end SAE work argues that standard SAEs may learn more about dataset structure than about computational structure, motivating KL-based functional alignment to the original model outputs. Conversely, FLINT and BM1.0 explicitly incorporate semantic interpretability criteria external to reconstruction: fidelity and visualization in FLINT, and BI-RADS–style mapping in radiomics. A plausible implication is that DDFID methods vary along at least two axes: functional faithfulness to the model and semantic accessibility to domain experts (Braun et al., 2024, Parekh et al., 2020).

Several limitations are explicitly recorded. In diffusion models, PCA whitening was needed to overcome anisotropic residual-stream variance, and ITDA, while very fast to build, yielded coarser and more abstract atoms than SAE. In end-to-end sparse dictionary learning, compute cost is materially higher than local SAE, with local SAE taking approximately 4–5 hours on one A100 and e2e variants approximately 11–12 hours. In DILA, thresholding the dictionary-to-code matrix to infer dependency relations is presented as a practical choice rather than a theoretically unique rule. In BM1.0, the DDFID interpretations are qualitative and no hard numeric feature thresholds were reported. In PIDF, the available material does not expose the equations or assembly procedure needed to determine whether its three-metric feature semantics should be regarded as a DDFID in the narrower dictionary-learning sense (Shabalin et al., 30 May 2025, Braun et al., 2024, Wu et al., 2024, Gorji et al., 21 Jul 2025, Westphal et al., 2024).

Taken together, these works place DDFID at the intersection of sparse representation learning, mechanistic interpretability, attribution methods, and domain translation. What remains stable across the variants is the attempt to move from raw learned features to a structured dictionary in which individual entries can be inspected, named, related to outputs, and in some settings manipulated. The differences among implementations indicate that DDFID is presently a research program and design pattern, not yet a single canonical methodology.