Concept bottleneck models are image classifiers that first predict human-readable attributes, such as "red wing" or "hooked beak," and then predict the label only from those attributes. Their main appeal is that a person can correct a wrong attribute at test time and watch the prediction change. They also have two known failures. When the attributes are incomplete, the model loses accuracy. When the attributes are soft numbers, the network hides extra information in them, which makes corrections unreliable. A common fix is to add a free "residual" channel that carries whatever the concepts miss. Done naively, this channel simply relearns the concepts and quietly overrides any human correction. This note proposes a residual channel that is statistically decorrelated from the concept predictions, not merely geometrically orthogonal to them. It also proposes an analytical gradient projection, borrowed from continual learning, that prevents task-driven fine-tuning of the backbone from changing what the concept head reads.
The construction touches only a pooled feature vector, so it applies unchanged to convolutional networks, vision transformers, and CLIP encoders. That property lets us test a question the architecture-dependence literature raises but rarely measures: does the reliability of concept interventions depend on the backbone, and can a training rule remove that dependence?
Core Claim
Recent work on concept bottlenecks has asked for orthogonality in the wrong place. HybridCBM adds a soft orthogonality loss between learned and LLM-defined concept vectors (Hybrid Concept Bottleneck Models). The obvious next step at this intersection would be to split the feature space into a concept subspace and its null space and route residual information through the null space. That step fails for a simple reason: two orthogonal directions in feature space can carry strongly correlated signals. If the backbone's features are correlated, the null-space residual still predicts the concepts, and the task head can still use it to undo a human correction.
What intervention faithfulness actually requires is that the residual carry no linear information about the concept predictions over the data distribution. With a linear task head, this condition has a clean consequence, by the standard partialling-out argument for linear regression: the coefficients on the concepts equal those of a concept-only model. A correction to a concept therefore moves the logit exactly as it would in an interpretable concept-only model, while the residual adds accuracy on top.
We claim that this decorrelation, together with a first-order guarantee that task gradients cannot rewrite the concept readout, recovers most of the gap to black-box accuracy without the intervention failures documented for leaky models. We also predict that intervention faithfulness becomes roughly invariant across backbones, whereas leaky joint training makes it backbone-dependent.
Research Program
Let $h \in \mathbb{R}^d$ be the pooled backbone feature and $c = W_c h \in \mathbb{R}^k$ the concept logits from a linear concept head. The concept head is trained only with concept supervision; it receives no task gradient, which closes the joint-training leak described by Concept Bottleneck Models and There Was Never a Bottleneck in Concept Bottleneck Models. The residual is the part of $h$ that is not linearly predictable from $c$:
$$r = Q\left(h - \Sigma_{hc}\Sigma_{cc}^{-1}(c - \mu_c)\right),$$
Here $\Sigma_{hc}$ and $\Sigma_{cc}$ are running estimates of the feature–concept cross-covariance and the concept covariance, $\mu_c$ is the running concept mean, and $Q$ is a learned low-rank map. The task head is $y = \mathrm{softmax}(Ac + Br)$. By construction, the cross-covariance of $r$ and $c$ is zero, so an intervention on $c$ cannot be "explained away" by $r$ at the linear level.
The second ingredient addresses the backbone. When the backbone is fine-tuned by the task loss flowing through $r$, those updates could still shift the features the concept head reads, which would degrade concepts silently. We therefore project the task-loss gradient at the feature layer onto the null space of the concept readout before backpropagating it:
$$g_h \leftarrow \left(I - W_c^\top (W_c W_c^\top)^{-1} W_c\right) g_h .$$
This guarantees that, to first order, task-driven updates leave $W_c h$ unchanged; only concept supervision can move the concepts. The projection is a $k \times d$ operation with $k=112$ on CUB. It therefore avoids the large projector memory that Diversifying the Mixture-of-Experts Representation for Language Models with Orthogonal Optimizer reports (about 1.38x). It also avoids the null-space exhaustion that paper describes, because we protect one $k$-dimensional readout rather than many growing subspaces.
The method borrows from three lines of work. Hard projection of updates onto the complement of protected activations comes from Continual Learning in Low-rank Orthogonal Subspaces and InfLoRA. The demonstration that analytical null-space projection beats soft suppression for concept editing comes from OrthoEraser, which reports 5 residual nudity instances on I2P against 121 for ESD. All three protect old tasks or benign features. None protects a human-intervenable concept readout inside a predictive model.
Concepts come from two sources. The first is the 112 ground-truth CUB attributes. The second is SAE-discovered concepts produced with the released M-CBM pipeline (Learning Concept Bottleneck Models from Mechanistic Explanations, code). This lets us test whether the residual rescues accuracy specifically when concepts are incomplete.
The architecture study runs the same head on ResNet-50, VGG19, ViT-B/16, and CLIP ViT-B/16. The motivation comes from the architecture-dependence literature. CAM-style explanations need global pooling (Zhang et al.), attention-based explanations do not transfer across model families (Early et al.), and residual-stream decompositions assume transformers (Gandelsman). Our head assumes only a feature vector. We do not claim the implementation invariance that Flow-based Attribution in Graphical Models formalizes, since the head still depends on $h$. We test something weaker and practical: whether intervention behavior agrees across backbones. Towards Universality found that transformers and Mamba share SAE features (mean max-pairwise correlation 0.681 against a 0.725 seed skyline). That result motivates expecting concept readouts to be comparable across backbones once leakage is removed.
Baselines
The naive reference points are the standard end-to-end model and a joint CBM. Concept Bottleneck Models reports CUB task error of 17.5% for the standard model and 19.9% for the joint CBM. It also reports roughly 48–49% error for CBMs against 62.7% for the standard model on CUB+Places. We retrain both on each backbone with the same augmentation, epochs, and a fixed hyperparameter search budget of 20 trials per method, tuned on a held-out split. The black-box accuracy on each backbone is the ceiling we aim to approach.
The strongest competitor on accuracy with interpretability is a CBM plus an unconstrained residual (side) channel trained jointly. Concept embedding models, discussed in Espinosa Zarlenga et al., are a close relative, though no reported number was found on Lacuna for them on CUB. These baselines should match our accuracy, because they have strictly more freedom. We beat them only if we match their accuracy within one point while showing intervention gains comparable to a concept-only model where they do not.
The strongest competitor on faithfulness is MCBM from There Was Never a Bottleneck in Concept Bottleneck Models. It removes leakage with a variational bottleneck on $I(Z_j;X\mid C_j)$; no reported CUB accuracy was found on Lacuna for it, so we reimplement it. For the incomplete-concept regime, the strongest competitor is M-CBM, which reports 73.70% CUB accuracy at five contributing concepts and 90.04% concept ROC-AUC (De Santis et al.). We run its released code on our backbones. HybridCBM, with its soft orthogonality loss, reports 68.38% on CIFAR-100 against 70.1% for a CLIP linear probe. We compare against it on CIFAR-100 with CLIP features.
Four ablations isolate the ingredients. The first is a residual restricted to the geometric null space of $W_c$ rather than the decorrelated residual; this tests the central claim that orthogonality is not enough. The second keeps the decorrelated residual but drops the gradient projection. The third replaces the projection with a soft penalty on changes to $W_c h$. The fourth freezes the backbone and decorrelates post hoc once.
Why not just…?
Why not just make the concepts binary, as hard CBMs do? Hard concepts stop decimal-place leakage (Havasi et al.), but they cannot recover accuracy when the concept list is incomplete. They also need straight-through or sampled gradients. We include a hard CBM with an added residual. Its weakness is that the residual still re-encodes the concepts, which is exactly the failure decorrelation targets. The ablation measuring intervention gain with and without decorrelation settles whether binarization adds anything once the residual is decorrelated.
Why not just add a soft decorrelation or orthogonality penalty? This is the cheapest competitor, and on a frozen backbone it may well match us. Soft penalties leave residual correlation that grows with the penalty's slack. Evidence from concept editing (OrthoEraser) and continual learning (Chaudhry et al.) shows that hard projection outperforms soft constraints when parameters keep moving. We sweep the penalty weight across the full tuning budget and report the best point on the accuracy–faithfulness curve. If any setting matches ours, we will report that the soft version suffices.
Why not just freeze the backbone and decorrelate once after training? This is the most serious threat. With a frozen backbone, the gradient projection does nothing and the method reduces to post-hoc linear partialling-out. Our claim for the projection applies only when fine-tuning the backbone is needed for accuracy. We test it by comparing frozen-plus-post-hoc against fine-tuned-plus-projection. If fine-tuning brings no accuracy gain on these benchmarks, we drop the projection and present the decorrelated residual alone as the contribution.
Why not just use post-hoc explainers such as InputIBA or Shapley-style attribution across architectures? These methods are architecture-agnostic. InputIBA reports an effective heat ratio of 0.476 on ImageNet against 0.283 for GradCAM (Zhang et al.). However, they explain without letting a user correct the model's reasoning. The quantity we measure is intervention behavior, which post-hoc maps do not provide.
Proposed Experiments
The main benchmark is CUB-200-2011 with its 112 attributes, on all four backbones. Across all experiments we report four measures:
- task accuracy;
- intervention gain, defined as accuracy as a function of the number of concepts corrected to ground truth, ordered by concept uncertainty;
- intervention consistency, defined as the correlation between the change in the full model's logit under a correction and the change in a concept-only model trained on the same concepts;
- residual–concept leakage, defined as the accuracy of a probe that predicts ground-truth concepts from $r$.
The decisive check has three parts. Our method must land within two points of black-box accuracy on each backbone. Its intervention consistency must be near that of the concept-only model. The geometric-null-space ablation must show low consistency and high probe leakage. If the null-space variant performs as well, the central claim is wrong.
Second, we run synthetic leakage tests on Shapes3D and MPI3D, where nuisance factors such as floor color and orientation are known. We measure how well those factors can be predicted from concept logits and from the residual. The concept logits should match MCBM's purity, while the residual is free to carry nuisance factors that matter for the task.
Third, we test incomplete concepts by dropping random subsets of 25%, 50%, and 75% of the CUB attributes. The prediction is that the advantage over MCBM grows as concepts are removed, because MCBM pays the honest accuracy drop and our residual does not. We repeat this with SAE concepts from M-CBM.
Fourth, under distribution shift on CUB+Places, we check whether interventions still correct background-induced errors. Koh et al.'s 48–49% CBM error is the reference point.
Fifth, for the architecture question, we report the spread of intervention consistency across the four backbones for joint CBM, the residual baseline, and our method. We predict a markedly smaller spread for our method. If the spread stays the same, the architecture-invariance claim fails and we will say so.
A secondary check on the OAI knee X-ray data, where correcting two concepts cut error from over 0.4 to about 0.3 in Koh et al., tests whether the method preserves intervention gains in a clinical regression setting. This check depends on obtaining data access.
References
- Hybrid Concept Bottleneck Models
- Concept Bottleneck Models
- There Was Never a Bottleneck in Concept Bottleneck Models
- Diversifying the Mixture-of-Experts Representation for Language Models with Orthogonal Optimizer
- Continual Learning in Low-rank Orthogonal Subspaces
- InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning
- OrthoEraser: Coupled-Neuron Orthogonal Projection for Concept Erasure
- Learning Concept Bottleneck Models from Mechanistic Explanations
- Exploiting Class Activation Value for Partial-Label Learning
- Model Agnostic Interpretability for Multiple Instance Learning
- Interpreting CLIP's Image Representation via Text-Based Decomposition
- Flow-based Attribution in Graphical Models: A Recursive Shapley Approach
- Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures
- Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts
- Addressing Leakage in Concept Bottleneck Models
- Fine-Grained Neural Network Explanation by Identifying Input Features with Predictive Information