The method, called Equation Discovery for Classification, or EDC, was developed by Guus Toussaint and Arno Knobbe of Leiden University. It extends a research tradition that has long flourished in regression: symbolic regression, the automated search for analytical formulas that fit numerical data. Theauthors of the study point out that equation discovery has historically been applied almost exclusively to problems where the target is a continuous number, such as recovering physical laws like the relationship between the pressure, volume, and temperature of a gas. Their contribution is to redirect that machinery toward binary classification, where the goal is to separate two classes cleanly and explainably.

The core idea is elegantly simple in concept. Rather than learning an opaque scoring function, EDC searches for a concise mathematical expression f(x) and a threshold theta, such that a data point is assigned to the positive class whenever f(x) meets or exceeds theta. For a linear equation, this is essentially the geometry underlying logistic regression or a linear support vector machine: a hyperplane slicing the feature space into two halves. What sets EDC apart is that the search is not confined to straight lines. By allowing nonlinear building blocks such as products of features and exponential terms, the algorithm can trace curved, interaction-driven boundaries while still producing an expression short enough for a domain expert to read, critique, and even correct by hand.

Technically, the framework rests on two pillars: a structured search and a dedicated numerical optimizer. The search space of candidate equations is defined by a configurable context-free grammar, a design choice inherited from classic work on declarative bias in equation discovery. The grammar constrains equations to sums of simple summands, including linear terms, products of two features, and exponential expressions, each parameterized by constants. Crucially, the grammar is redundancy-aware: constructions that would produce syntactically different but semantically identical equations are pruned in advance, since, for example, subtraction between summands is unnecessary when a constant coefficient can simply be negative. Traversing this space exhaustively is impossible for all but the smallest problems, so the algorithm employs beam search, iteratively refining the most promising candidate equations level by level while keeping only a fixed number of survivors at each stage.

Fitting the constants inside each candidate equation proved surprisingly subtle. Although every equation in the grammar is differentiable, gradient descent performed poorly in practice, largely because of the exponential terms that pepper the search space. The authors instead adopted a tailored hill-climbing procedure that first samples a large pool of random constant configurations, then concentrates its remaining budget on refining the best ones. In a systematic comparison against off-the-shelf optimizers from the SciPy library, including Powell, Cobyqa, Cobyla, Nelder-Mead, and stochastic gradient descent, this hill climber achieved the highest mean area under the ROC curve across one hundred randomly generated test problems. Perhaps most strikingly, a simple random-sampling baseline already reached a mean AUC of 0.9974 with a budget of one thousand evaluations, revealing that the inner optimization problem is more tractable than one might fear.

The experiments on synthetic data produced one of the study’s most intriguing findings. When Gaussian noise was injected into datasets whose generating decision boundaries were known, EDC did not merely match the original boundary; beyond a certain noise level it outperformed it. The explanation is a phenomenon the authors describe carefully: noise pushes data points near a curved boundary across it, and concave sections of that boundary collect more stray points than they lose, causing the effective boundary embedded in the data to drift away from its original position and gradually straighten. EDC, fitting the data rather than the hidden formula, tracks this shifted, smoothed boundary, achieving higher AUC than the very equation that generated the data. As noise grew from negligible to substantial across seventeen hundred artificial datasets, even a restricted linear version of EDC eventually beat the original nonlinear boundary.

The framework also proved capable of reconstructing notoriously difficult structures, including XOR-like and interaction-driven boundaries, and of fitting data produced by Gaussian clusters where no explicit target equation exists at all. On these cluster-based problems, which the authors describe as closest to real-world conditions among their artificial experiments, EDC outperformed existing symbolic classification approaches while landing near state-of-the-art black-box methods such as random forests, multi-layer perceptrons, and radial-basis-function support vector machines, all of which posted mean AUC values around 0.97.

Real-world benchmarks reinforced the pattern. Across nine binary classification datasets from the UCI repository, spanning tasks from banknote authentication to income prediction, EDC achieved a higher AUC than every competing equation-discovery-based classifier on every dataset, and beat simple decision trees across the board. A critical distance analysis showed that random forests, neural networks, and SVMs did not statistically significantly outperform EDC, even though those black-box methods won on several datasets, particularly ionosphere and sonar, where feature-class relationships appear to lie outside EDC’s current set of building blocks. On the Adult income dataset, the discovered equation offered a vivid demonstration of interpretability in action: one term acted as a penalty on years of education only when the individual appeared as a child in the household, while an exponential term over marital status effectively functioned as an if-else statement, adding roughly 28,586 to the score for married individuals living with a spouse and a negligible 8 otherwise.

The method’s main weakness is computational cost. In its default configuration, with a search depth of six and a beam width of ten, EDC takes dramatically longer than the sub-second runtimes of conventional classifiers, largely because pairwise interaction terms grow quadratically with the number of features and because one-hot encoding of categorical variables inflates the feature count. The authors show, however, that the expense is largely optional. Restricting the search depth, narrowing the beam, or replacing interaction terms with simple quadratic terms produced speed-ups of up to a factor of fifty at only a marginal loss of accuracy, with no statistically significant difference in performance across the benchmark suite. The authors also note that the framework is not limited to two classes: standard one-versus-rest schemes can extend EDC to multi-class problems, with each of the resulting equations explaining when a particular class label prevails.

The work arrives amid a broader movement toward explainable machine learning, driven by domains such as medicine, finance, and criminal justice where decisions carry real consequences and regulators increasingly demand justification. EDC offers a principled bridge between symbolic regression and classification, delivering models that are simultaneously compact, expressive, and auditable. While it may not dethrone random forests or deep networks on raw accuracy, it demonstrates that the gap is small enough, and the interpretability dividend large enough, that equations once again deserve a seat at the machine learning table. The authors have released their code, experimental setup, and datasets openly so that the entire study can be reproduced and the grammar extended by practitioners in new domains.

Subject of Research: Equation discovery for binary classification using interpretable symbolic decision boundaries

Article Title: Equation Discovery for Classification: Finding Interpretable Symbolic Specifications of the Decision Boundary

Article References: Toussaint, G., & Knobbe, A. (2026). Equation Discovery for Classification: Finding Interpretable Symbolic Specifications of the Decision Boundary. Machine Learning, 115(9), Article 215. https://doi.org/10.1007/s10994-026-07156-1

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07156-1

Keywords: equation discovery, binary classification, symbolic regression, decision boundary, interpretable machine learning, beam search, symbolic classification, explainable AI, context-free grammar, UCI datasets, machine learning, Equation

Traditional concept bottleneck models (CBMs) have long been employed as a beacon for enhancing interpretability in AI systems. These models impose an intermediate representation—“concepts”—on the path to final prediction decisions. Such concepts, grounded in human language or domain expertise, provide a structured explanation: for example, a model identifying a bird species might pinpoint features like “yellow legs” or “blue wings” before delivering its classification. This intermediate step acts as a conceptual bottleneck, theoretically allowing users to peer into the model’s “thought process” and verify the factors influencing its conclusion.

However, the utility of classic CBMs is hampered by a fundamental limitation: the concepts are typically predefined by human experts or large language models, and inherently may not align perfectly with the complexities or nuances of the specific task or dataset. This mismatch can degrade both the accuracy of predictions and the fidelity of explanations. Furthermore, models often suffer from “information leakage,” where latent knowledge not captured by explicit concepts influences predictions surreptitiously, impairing transparency and trustworthiness. The result is a paradoxical situation: the AI might use relevant but obscured information outside the intended explanatory framework.

Confronting this issue head-on, MIT’s new methodology departs from conventional reliance on externally imposed concepts. Instead, it leverages the deep learning model’s existing internal knowledge. Since advanced computer vision models are typically trained on vast, diverse datasets, they inherently learn an abundance of latent features representing intricate patterns and discriminative information relevant to the task. The novel technique taps into this reservoir to distill meaningful, task-specific concepts that the original model has effectively “discovered” on its own.

The process begins with a specialized deep learning architecture known as a sparse autoencoder, a network designed to compress and then reconstruct data while isolating the most salient features. By applying this autoencoder to the target model’s learned representations, the researchers selectively extract a concise set of meaningful features that encapsulate essential discriminatory information. These distilled features are effectively the raw “concepts” embedded in the original model’s knowledge.

Next, a cutting-edge multimodal large language model (LLM) is employed to translate these distilled, abstract features into comprehensible plain-language descriptions. This step is crucial; it renders the otherwise inscrutable feature vectors into semantic concepts accessible to humans, enabling precise annotation and interpretation. Using this annotated data, the team trains a concept bottleneck module capable of identifying the presence or absence of each concept within individual images, thereby anchoring the model’s explanatory framework directly to its inherent learned knowledge.

Incorporating this concept bottleneck module back into the original computer vision model creates a powerful synergy: predictions are compelled to rely solely on the extracted learned concepts. This integration not only preserves the model’s high predictive power but fundamentally enhances interpretability by forcing a transparent, concept-based reasoning process. Consequently, medical professionals, researchers, or end-users can query the model’s decision pathway in terms intelligible to their expertise, bridging the gap between opaque AI predictions and actionable understanding.

One of the significant innovations in this methodology is the deliberate limitation imposed on the number of concepts utilized per prediction. By constraining the model to select just five concepts, the researchers ensure that explanations remain succinct, focused, and comprehensible rather than overwhelmed by an unmanageable multitude of factors. This also functions as a rigorous filter, compelling the system to prioritize the concepts most relevant to each specific instance—a crucial feature for practical high-stakes applications like diagnosing skin lesions or species classification.

In rigorous evaluations comparing this new approach against state-of-the-art concept bottleneck models, the MIT team demonstrates superior accuracy alongside enhanced explanatory clarity. Testing on challenging datasets, including those for bird species identification and dermatological image classification, their method not only matches but frequently surpasses performance benchmarks while generating more precise, conceptually relevant explanations. Such improvements signify a notable stride toward reconciling the historically difficult trade-off between interpretability and performance in AI models.

Despite these advances, the researchers acknowledge ongoing challenges, particularly regarding the persistence of some degree of information leakage and the inherent complexity of fully interpretable AI. While their approach markedly reduces the risk of undisclosed concepts influencing predictions, absolute elimination remains elusive. Future work is poised to investigate multi-layered concept bottlenecks to more effectively seal off unwanted information pathways and enhance robustness against leakage.

Scaling the approach also promises exciting avenues for growth. By deploying larger, more capable multimodal LLMs for concept annotation and leveraging expanded training datasets, the researchers aim to further boost both the fidelity of explanations and the predictive prowess of concept-driven models. These enhancements could broaden applicability across diverse domains and spur widespread adoption in critical AI-powered decision systems.

The implications of this research extend far beyond academic curiosity. In clinical contexts, for example, transparent AI tools can provide clinicians with justifiable evidence when interpreting medical images, fostering informed decision-making and bolstering patient trust. More broadly, improved accountability in AI systems bridges a crucial ethical gap, addressing societal concerns about opaque “black-box” models and contributing to safer, fairer, and more reliable artificial intelligence technologies.

The collaboration underlying this advancement brought together international expertise, featuring contributions from Antonio De Santis, a graduate student at Polytechnic University of Milan and CSAIL visiting scholar, alongside colleagues Schrasing Tong, Marco Brambilla, and CSAIL principal researcher Lalana Kagal. Their work, recently accepted for presentation at the International Conference on Learning Representations, represents a milestone in concept-driven AI interpretability research.

In summary, MIT’s innovative methodology charts a promising course toward AI models that do not merely compute predictions but elucidate their reasoning through human-understandable concepts inherently learned during training. By extracting and harnessing these latent knowledge structures, this approach synthesizes accuracy with interpretability, promising a future where AI transparency is not an afterthought but a foundational feature integral to systems that impact lives and society at large.

Subject of Research: Explainable Artificial Intelligence, Concept Bottleneck Models, Computer Vision, Machine Learning Interpretability

Article Title: Extracting Learned Concepts for Enhanced Explainability in Computer Vision Models

News Publication Date: Not specified in the source

Web References: Research Paper on OpenReview

Keywords

Artificial Intelligence, Explainability, Concept Bottleneck Modeling, Computer Vision, Machine Learning, Interpretability, Sparse Autoencoder, Large Language Models, Medical Diagnostics, Black-box Models, Information Leakage, Multimodal Models