Correlation-Aware Reweighting for Concept Bottleneck Heads

Concept bottleneck models are image classifiers that must explain themselves. They first score an image against a list of human-readable concepts ("long, blue-gray wings", "yellow breast") and then make the final decision with a linear layer over those scores. Modern versions get the concept list from a language model and score it with CLIP, so the list is large and full of near-duplicates. A sparse final layer is supposed to keep each explanation short. On a list full of synonyms, though, the standard sparsity penalty picks one member of each synonym cluster more or less at random, or keeps several of them. The resulting explanation looks sparse, but it changes from one training run to the next, and some of its slots are wasted on repeating the same idea.

We propose to replace that penalty with a reweighting rule from the empirical-Bayes view of sparse regression. Under this rule, the penalty on each concept grows when the concepts already in use can explain it. A paraphrase of a selected concept then becomes expensive to add, and a concept with genuinely new information stays cheap. The claim we will test is narrow. On LLM-generated concept libraries, this correlation-aware reweighting should give explanations that are shorter at equal accuracy and more stable across retraining than uniform $\ell_1$, magnitude-based reweighting, or deduplicating the list beforehand.

Core Claim

The obvious idea at this intersection is to apply Candès-style reweighted $\ell_1$, the "penalize small weights more" rule, to the concept head in order to undo the magnitude bias of the Lasso. The research file notes this has not been tried. Sparse Concept Bottleneck Models: Gumbel Tricks in Contrastive Learning compared only static $\ell_1$, a contrastive loss and Gumbel-Softmax selection. We think magnitude bias is the wrong target. A reviewer would rightly point out that "select with $\ell_1$, then refit the chosen weights without a penalty" removes magnitude bias almost for free. Refitting does nothing about which concepts get selected, however, and selection is what an explanation actually is.

The problem that is specific to concept libraries is correlation. Near-synonymous text descriptions make the concept-score matrix badly conditioned, which violates the conditions under which $\ell_1$ provably finds the right sparse set. Dual-Space Analysis of the Sparse Linear Model shows that empirical-Bayes (Type II) estimation, written as a reweighted $\ell_1$ scheme, still recovers sparse solutions under clustered, correlated dictionaries. It can do this because its weights depend on the dictionary geometry and not only on coefficient magnitudes.

We therefore claim that porting these geometry-dependent weights to a multiclass concept head is the ingredient that matters. A magnitude-only reweighting control is included to isolate it.

Method

Let $\Phi \in \mathbb{R}^{n \times K}$ hold frozen CLIP concept scores for $n$ images and $K$ concepts, with column $\phi_j$ for concept $j$. Let $B \in \mathbb{R}^{K \times C}$ be the head weights for $C$ classes, with row $b_j$. We penalize whole rows, so a concept is either used for every class or for none. This makes "number of concepts used" a meaningful count, following the joint-sparsity reweighting of Recovery Algorithms for Vector-Valued Data with Joint Sparsity Constraints. Each outer iteration solves

$$\min_B \; \mathcal{L}(\Phi B, y) + \lambda \sum_{j} w_j \|b_j\|_2,$$

where $\mathcal{L}$ is the multinomial logistic loss and $\lambda$ is a single global strength. After each solve, the weights are updated with the dual-space rule

$$w_j \leftarrow \Big[\phi_j^\top \big(\lambda I + \Phi\, \Gamma\, \Phi^\top\big)^{-1} \phi_j\Big]^{1/2}, \qquad \Gamma = \mathrm{diag}\big(\|b_k\|_2 / w_k\big).$$

Here $\Gamma$ collects the current per-concept variance estimates. The quadratic form is small when $\phi_j$ already lies in the span of the active concepts, which raises $w_j$ for redundant concepts. This is the behavior the method relies on.

Two practical points follow. First, the logistic loss is not Gaussian, and Dual-Space Analysis of the Sparse Linear Model flags that this breaks exact marginalization. We take the standard route of a local quadratic (iteratively reweighted least squares) approximation at each outer step. We do not claim the theoretical guarantees carry over; we test whether the behavior does. Second, the $n \times n$ inverse is computed in $K$-dimensional form through the matrix inversion identity. This keeps a weight update to one $K \times K$ solve per iteration, which is cheap even at $K = 4{,}505$ for ImageNet.

The magnitude-only control, from Enhancing Sparsity by Reweighted L1 Minimization, instead uses $w_j = 1/(\|b_j\|_2 + \epsilon)$. That paper also reports the known sensitivity of this control to the choice of $\epsilon$. The dual-space rule has no $\epsilon$, a secondary advantage we will check.

Research Program

We will build three pieces. The first is a drop-in head that runs the correlation-aware reweighted group penalty on frozen concept scores, compatible with the Label-free CBM and Concept Discovery Model pipelines. The concept-score extraction will reuse the released ConceptDiscoveryModels code so that every method sees identical inputs.

The second is a controlled synonym-injection protocol. For each base concept, we ask a language model for $r$ paraphrases, with $r$ swept from 0 to 4, and append them to the library. This gives a known ground-truth cluster structure to grade against.

The third is a set of evaluation measures aimed at explanation quality rather than accuracy alone. Accuracy is reported as a curve against the number of distinct concepts used. Selection stability is the Jaccard overlap of selected concept sets across bootstrap resamples and seeds. The synonym-collapse rate is the fraction of selected concepts that share a paraphrase cluster with another selected concept. Per-decision simplicity is measured with the Sparse Explanation Value from Improving Decision Sparsity, the number of concept changes needed to flip a decision. That paper shows global sparsity does not guarantee short local explanations, so we measure both rather than assume one implies the other.

Baselines

The naive reference point is a dense multinomial logistic head over all concepts, tuned only for accuracy. It sets the accuracy ceiling for frozen concept scores and shows how many concepts an unconstrained head spreads its weight over.

The standard sparse head is the elastic-net or $\ell_1$ linear layer used in Label-free Concept Bottleneck Models. That model maps CLIP features to GPT-generated concepts, filters the list by text-embedding similarity, and fits a sparse final layer. It reports 71.95% top-1 on ImageNet and 86.4% on CIFAR-10. We run it both with and without its text-similarity filter. Its "with filter" variant is the direct competitor to our claim, because it attacks redundancy before training rather than during it.

We include $\ell_1$ selection followed by an unpenalized refit on the chosen concepts. It is the cheapest fix for magnitude bias, and if it matches us, our story about selection is wrong.

We include the two strongest learned-selection methods. Sparse Concept Bottleneck Models uses Gumbel-Softmax selection and reports 80.02% accuracy on CUB-200. Sparse Linear Concept Discovery Models learns per-image binary concept indicators under a fixed variational prior. It reports 95.30% on CIFAR-10 with a ViT-B/16 backbone and keeps 6.96% of 4,505 concepts on ImageNet. For CIFAR-100 and Places365, no reported result was found on Lacuna for any of these methods, so we will produce all numbers on those datasets ourselves. The 84.1% CUB result of Auxiliary Losses for Learning Generalizable Concept-based Models uses human concept annotations and a trained backbone, so it belongs to a different regime and is cited only for context.

The key ablation is magnitude-only reweighting with identical row grouping, solver and outer-iteration count. It changes exactly one thing, whether the weights see the concept geometry.

All methods use the same frozen backbone and concept scores, the same train/validation splits, and the same budget of 30 validation-tuned configurations. Each is reported along a sweep of sparsity strength, so comparisons are made at matched numbers of concepts and not at single points. "Beating" a baseline means one of two things. We either match its accuracy within one standard deviation across five seeds while using fewer distinct concepts and showing higher selection stability, or we show strictly higher accuracy at every matched concept budget. Reproduced numbers that miss the cited figures by more than two points will be flagged rather than silently used.

Why not just…?

Why not just use $\ell_1$ and refit the selected weights? Refitting repairs the shrinkage of large weights. It cannot change which synonym was picked, so it should do nothing for stability or synonym collapse. That is a falsifiable prediction. If refit-$\ell_1$ matches us on the synonym-injection sweep, the geometry-aware weights are unnecessary and we will say so.

Why not just deduplicate the concept list by text similarity first? This is the most serious threat, because it is cheap and already standard in Label-free Concept Bottleneck Models. Our argument is that redundancy that matters for the head is visual, in the image-score matrix, not textual. Two descriptions can read differently yet fire on the same images, and two similar-sounding ones can differ visually. The experiment that settles it measures the visual correlation of concept pairs that survive the text filter. We then compare filter-then-$\ell_1$ against our method at matched concept counts. If the filter closes the gap, the proposal reduces to a better filter, which is still a result but a smaller one.

Why not just use Gumbel-Softmax selection, which already beat $\ell_1$? It beat static $\ell_1$ in Sparse Concept Bottleneck Models, and it may well beat us on raw accuracy. Its selection is stochastic and depends on a temperature schedule. Nothing in that paper addresses how it behaves on correlated libraries, so we expect lower run-to-run stability under synonym injection. The comparison at matched concept budgets will show whether that expectation holds. A loss on both accuracy and stability would falsify our claim.

Why not just magnitude-based reweighting? Its weights are blind to correlation. Two equally strong synonyms receive equal, small penalties and can both survive. The magnitude-only ablation tests exactly this.

Proposed Experiments

The first decisive check is synthetic. We plant a sparse linear head over CLIP concept scores on CUB-200, inject paraphrases at increasing $r$, and measure whether each method recovers one concept per planted cluster. The prediction is that synonym collapse grows with $r$ for $\ell_1$, refit-$\ell_1$ and magnitude reweighting, but stays near zero for the dual-space rule. Accuracy at a fixed concept budget should stay flat for the dual-space rule and degrade for the others.

The second check is on real libraries for CUB-200, CIFAR-10, CIFAR-100, Places365 and ImageNet. We plot accuracy against distinct concepts used and report bootstrap Jaccard stability and mean Sparse Explanation Value per decision. The claim succeeds if the dual-space curve lies above the $\ell_1$ and magnitude-reweighting curves at budgets of 20–200 concepts. It must also come within one point of the Gumbel and Concept Discovery results at their reported operating points while showing higher stability.

The third check is about cost. We report wall-clock time per outer iteration on ImageNet, since A Unified DNN Weight Compression Framework Using Reweighted Optimization Methods identifies repeated retraining as reweighting's practical weakness. The method fails this check if it needs more than roughly ten outer solves to stabilize.