KANFIS: A Neuro-Symbolic Framework for Interpretable and Uncertainty-Aware Learning

arXiv:2602.03034 · cs.AI · Submitted 2026-02-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "KANFIS: A Neuro-Symbolic Framework for Interpretable and Uncertainty-Aware Learning".

Tom: Adaptive Neuro-Fuzzy Inference System (ANFIS) was designed to combine neural network learning with fuzzy logic reasoning, but conventional architectures suffer from structural complexity and rule explosion.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So, we’re looking at the title "KANFIS: A Neuro-Symbolic Framework for Interpretable and Uncertainty-Aware Learning" and who wrote it—Binbin Yong, Haoran Pei, Jun Shen, Haoran Li, Qingguo Zhou, Zhao Su. Basically, the core idea is creating a compact neuro-symbolic architecture that combines fuzzy reasoning with additive function decomposition to make things interpretable and aware of uncertainty.

Jane: That sounds like they are trying to bridge the gap between complex neural networks and transparent rule-based systems by using this new KANFIS framework. It’s not just about fitting data; it’s about making the learning process itself transparent.

Lu: The authors are tackling that fundamental trade-off where you usually have to choose between having a powerful nonlinear approximation and keeping the model simple enough to interpret, especially when dealing with complex features in high dimensions.

Meng: I'm thinking about how this impacts practical deployment; if the architecture is designed so rule complexity scales linearly with input dimension, that means we can actually build these models for bigger datasets without needing massive computational resources just to manage the rule set.

Lalam: That linear scaling idea is really compelling because it suggests a path toward AI systems that are both expressive and manageable in real-world scenarios.

The paper's summary: Tom: So, what they’re summarizing is how KANFIS integrates the symbolic reasoning of ANFIS with the additivity theorem from Kolmogorov-Arnold networks by making the network edges act as learnable membership functions that are aggregated in an additive way.

Jane: That sounds like a clever way to structure things; they are taking the best parts of both worlds and combining them through this specific aggregation mechanism. It’s essentially scaling the model parameters and rule complexity linearly with input dimensionality instead of exponentially, which is a big deal for scalability.

Lu: They also specifically mention that KANFIS supports both Type-one fuzzy logic systems and Interval Type-two fuzzy logic systems, which allows for the explicit modeling of uncertainty and ambiguity through upper and lower membership degrees defined by equations three and four in the paper.

Meng: That explicit handling of uncertainty is crucial; in industrial applications, knowing not just what the prediction is but also how uncertain we are about that prediction gives operators much more confidence when making decisions.

Lalam: It’s wonderful to see a framework that isn't afraid to model ambiguity directly rather than trying to smooth it out or ignore it, which really pushes the boundaries of what we expect from interpretable AI.

The paper's improvements: Tom: Moving on to how they improve things, they introduce several mechanisms. They use sparse masking mechanisms specifically to generate compact and structured rule sets for KANFIS, which is key for achieving intrinsic interpretability.

Jane: The introduction of entropy-based antecedent sparsity (Rsparse) is interesting because it actively forces the mask values toward zero or one, which effectively prunes unnecessary connections in the model, making it much more readable.

Lu: They also have rule distinctiveness regularization (Rdistinct), which penalizes overlap between the firing patterns of different rules by minimizing pairwise cosine similarity between them, ensuring each rule specializes in a unique region of the input domain.

Meng: The authors show that without these regularization terms, the rules tend to incorporate all available features, leading to dense and tangled logic; so these penalties are necessary to keep the model from becoming an unreadable mess as it grows.

Lalam: Those regularization terms really speak to the goal of creating a system that is not just accurate but also inherently structured and easy for humans to follow, which I think is essential for long-term adoption.

Conclusion: Tom: So, wrapping up the paper "KANFIS: A Neuro-Symbolic Framework for Interpretable and Uncertainty-Aware Learning," the main takeaway is that KANFIS achieves competitive performance against neural and neurofuzzy baselines while ensuring intrinsic interpretability through its additive aggregation and sparse masking.

Jane: It’s a lot to digest; they managed to unify fuzzy reasoning with additive function decomposition in a way that handles uncertainty explicitly, which is something we haven't seen implemented this cleanly before.

Lu: The paper highlights that the scaling of the Additive Fuzzy System, or AFS KANFIS, scales linearly with respect to the input dimension N, contrasting sharply with product fuzzy systems which scale exponentially; it shows a clear path toward practical scalability.

Meng: From an engineering standpoint, this linear scaling is what makes it viable for high-dimensional industrial data applications where rule explosion would otherwise make training impossible or the model too slow to run effectively.

Lalam: What I really see here is a future where AI models can provide not just a prediction, but a set of rules that are rigorously consistent with underlying physical laws, which opens up so many new possibilities for domain-specific applications.

School of Information Science and Engineering, Lanzhou University · School of Computing and Information Technology, University of Wollongong, Australia · Department of Data Science and Artificial Intelligence, Monash University

cs.AI

Submitted: 2026-02-03

Updated: 2026-09-28

Importance score: 90/100

The gist: Adaptive Neuro-Fuzzy Inference System (ANFIS) was designed to combine neural network learning with fuzzy logic reasoning, but conventional architectures suffer from structural complexity and rule

Key concepts

KANFIS
A compact neuro-symbolic architecture that merges fuzzy reasoning with the additivity theorem of Kolmogorov-Arnold networks. It designs network edges as learnable membership functions, scaling model complexity linearly with input dimensions instead of exponentially.
Type-1 and Interval Type-2 (IT2) Fuzzy Logic
The system supports both standard Type-1 fuzzy logic and Interval Type-2 systems. IT2 allows for the explicit modeling of uncertainty by defining membership degrees using both upper and lower bounds, providing a richer representation of ambiguity.
Soft Rule Antecedent
This is a measure calculated for each rule that quantifies how strongly an input feature activates the rule's fuzzy basis. It is found by summing the upper and lower membership degrees of the input feature across all associated fuzzy sets.
Entropy-based Antecedent Sparsity (Rsparse)
A regularization technique used to improve model readability. It minimizes the pixel-wise binary entropy of a soft-masking matrix, which forces mask values to be either zero or one, effectively pruning irrelevant connections and simplifying the logic.

Terminology

Summary

Adaptive Neuro-Fuzzy Inference System (ANFIS) was designed to combine neural network learning with fuzzy logic reasoning, but conventional architectures suffer from structural complexity and rule explosion. This work proposes the Kolmogorov-Arnold Neuro-Fuzzy Inference System (KANFIS), a compact neuro-symbolic architecture that unifies fuzzy reasoning with additive function decomposition to achieve competitive performance while ensuring intrinsic interpretability and uncertainty awareness.

How it works

  1. KANFIS integrates the symbolic reasoning of ANFIS with the additivity theorem of Kolmogorov-Arnold networks by designing network edges to function as learnable membership functions, which are then aggregated using an additive aggregation mechanism. This configuration scales both model parameters and rule complexity linearly with input dimensionality rather than exponentially.

  2. The architecture supports both Type-1 (T1) and Interval Type-2 (IT2) fuzzy logic systems, enabling the explicit modeling of uncertainty and ambiguity in fuzzy representations through upper and lower membership degrees, defined by Equations (3) and (4).

  3. For inference, the quantity measuring how strongly input feature xi activates the k-th fuzzy basis along an edge is calculated as a center-of-sets type-reduction: ϕi,j,k(xi) = UMFi,j,k(xi) + LMFi,j,k(xi).

  4. Each edge aggregates K basis memberships to form a soft rule antecedent: ei,j (xi) = X K k=1 ϕi,j,k(xi). This quantity serves as the soft rule antecedent, interpreting the interaction between input xi and hidden unit j as: “If feature xi belongs to one of the fuzzy sets associated with unit j, then the rule fires with strength ei,j (xi).”

  5. The hidden representation for a given unit j is formed through a Kolmogorov-style summation across all incoming fuzzy edges: hj = X din i=1 ei,j (xi). This aggregation acts as a learned nonlinear embedding constructed entirely through fuzzy functional components.

Structural Components and Theoretical Foundation

  1. KANFIS is fundamentally an additive network structure where the mapping from input x to output fKANFIS(x) is defined by fKANFIS(x) = X R j=1 wj · hj (x) + b. The core component, the edge function ϕij (xi), is a summation of M center type-reduced IT2 Gaussian MFs.

  2. Multiple KANFIS layers can be stacked to form a deep fuzzy inference hierarchy, where the transformation for the l-th layer is h(l) = F(l) h(l−1). To stabilize training and maintain numeric consistency across layers, a normalization is applied after each layer except the final one: h(l) ← Norm(h(l)), l < L.

  3. The final output yˆ = Wh (L) + b is obtained via a linear defuzzification layer, which aggregates all activated fuzzy rules into a single crisp prediction, acting as a generalized consequent layer of a Takagi-Sugeno fuzzy system.

Interpretability and Regularization

  1. To ensure model compactness and interpretability, the total objective function is formulated as Ltotal = Ltask + λsRsparse + λdRdistinct, balancing predictive accuracy with structural constraints.

  2. Entropy-based Antecedent Sparsity (Rsparse) is introduced to improve human readability by minimizing the pixel-wise binary entropy of a learnable soft-masking matrix M, which forces mask values towards 0 (irrelevant) or 1 (selected), thereby pruning unnecessary connections.

  3. Rule Distinctiveness (Rdistinct) penalizes the overlap between the firing patterns of different rules by minimizing the pairwise cosine similarity between all distinct pairs of rules, ensuring that each rule specializes in a unique sub-region of the input domain.

Performance and Scalability

  1. Empirical results demonstrate that KANFIS achieves competitive performance against representative neural and neurofuzzy baselines across various datasets, outperforming models like ANFIS in high-dimensional settings where standard grid partitioning becomes computationally intractable due to rule explosion.

  2. Ablation experiments confirm that the regularization term plays a crucial role in controlling model complexity, showing that without regularization, rules tend to incorporate all available features, leading to dense and entangled logic.

  3. The theoretical analysis shows a significant efficiency disparity: while Product Fuzzy Systems (PFS) scale exponentially with input dimension N (Pprod ∝ O(MN)), the Additive Fuzzy System (AFS) KANFIS scales linearly with respect to the input dimension N (Psum ∝ O(N · H · Pψ)), demonstrating superior practical scalability.

  4. Interpretability analysis on datasets like CCPP and MHR shows that extracted IF-THEN rules are rigorous and consistent with expert knowledge, successfully rediscovers underlying physical or physiological laws governing the data.

Universal Approximation Capability

Improvements for AI systems

Here are the specific improvements that can be made to existing AI systems by implementing the KANFIS framework, along with a description of what these improved systems will be able to do:

  1. Improve Interpretability and Trustworthiness in High-Stakes Systems (e.g., Autonomous Vehicles, Medical Diagnostics, Financial Risk Modeling).

  2. Enhance Robustness under Uncertainty by Explicitly Modeling Ambiguity and Aleatoric Uncertainty.

  3. Achieve Superior Scalability for High-Dimensional Data without the Curse of Dimensionality in Neuro-Fuzzy Architectures.

  4. Generate Concise, Human-Readable, and Domain-Consistent Symbolic Rules from Complex Data Patterns.

Specific Capabilities of the Improved KANFIS System:

  1. A system can perform complex predictions while providing a transparent If-Then rule set that maps directly to established physical or physiological laws (as demonstrated by the CCPP and MHR case studies).

  2. The system can explicitly quantify and visualize uncertainty (using Interval Type-2 fuzzy logic) rather than relying on point estimates, allowing operators to understand the confidence bounds of a decision.

  3. The system can efficiently process high-dimensional inputs (e.g., 100+ sensor readings) by scaling its parameter count linearly with the input dimension, avoiding the exponential rule explosion that plagues traditional ANFIS models.

  4. The system can automatically prune redundant features and rules using entropy-based masking and distinctiveness regularization, ensuring each learned rule focuses on a minimal, highly informative subset of inputs (e.g., reducing feature usage per rule by up to 63% on the Spambase dataset) to maximize human readability and domain relevance.

Sources

Related papers