A Sugeno Integral View of Binarized Neural Network Inference

Abstract: In this article, we establish a precise connection between binarized neural networks (BNNs) and Sugeno integrals. The advantage of the Sugeno integral is that it provides a framework for representing the importance of inputs and their interactions, while being equivalent to a set of if-then rules. For a hidden BNN neuron at inference time, we show that the activation threshold test can be written as a Sugeno integral on binary inputs. This yields an explicit set-function representation of each neuron decision, and an associated rule-based representation. We also provide a Sugeno-integral expression for the last-layer score. Finally, we discuss how the same framework can be adapted to support richer input interactions and how it can be extended beyond the binary case induced by binarized neural networks.

• The paper presents an exact derivation that recasts binarized neural network activation as a Sugeno integral, unifying thresholding with rule-based aggregation.
• It translates hidden and output layer computations into capacity-based evaluations, enabling enhanced explainability and formal verification.
• The framework lays groundwork for integrating qualitative reasoning in neural models and extends naturally to richer, quantized representations.

A Sugeno Integral Perspective on Binarized Neural Network Inference

Introduction

The paper "A Sugeno Integral View of Binarized Neural Network Inference" (2604.17967) establishes a structural correspondence between binarized neural networks (BNNs) and the Sugeno integral, a classical qualitative aggregation operator in multi-criteria decision-making and fuzzy logic. BNNs, with weights and activations constrained to bipolar values, provide a natural setting to directly connect neural thresholding to qualitative rule-based aggregation via capacities and focal sets. The authors present exact derivations translating hidden-layer and output-layer BNN inference into the Sugeno integral framework, yielding a capacity-based, rule-oriented representation for these models. The implications extend to explainability, verification, and neurosymbolic reasoning, while laying the groundwork for future developments in qualitative and quantized neural computations.

Background and Motivation

BNNs, as defined by Hubara et al., operate with weights and activations fixed to $\{-1, +1\}$, and their inference reduces to evaluating thresholded weighted sums. These properties have made BNNs attractive for resource-constrained inference, but also for program analysis, verification, and, increasingly, explainable AI. Meanwhile, the Sugeno integral is a parameterized aggregation operator defined through a monotone set function ("capacity") and is tightly linked to rule-based systems, minimal sufficient conditions ("focal sets"), and possibility theory.

Prior work focused on learning Sugeno measures with neural networks or using rule-based systems for reasoning, but the explicit, structural translation from BNN computations to the Sugeno integral—mirroring both the inference and the underlying rule structure—was not previously formalized as rigorously as in this paper.

Main Contributions

Sugeno Integral Representation of BNN Inference

The core result demonstrates that the threshold activation of a hidden-layer BNN neuron can be rewritten exactly as a Sugeno integral on a suitable $\{0,1\}$-binary, "polarized" encoding of the input. Each neuron, defined by a bipolar weight vector $\bar w\in\{-1, +1\}^n$ and a threshold $\theta$, is mapped to a capacity $\mu_{\bar w, \tau}$ on the literal space $I = \{1^+, 1^-, ..., n^+, n^-\}$, where the focal sets correspond to minimal sufficient input combinations activating the neuron (i.e., subsets of literals corresponding to matching the weight configuration and meeting the threshold).

The inference for a hidden neuron is then:

• Encode the input to $\{0,1\}^{2n}$ with "positive" and "negative" literals.
• Define $\Lambda(\bar w)$ as the subset of literals agreeing with the signs of $\bar w$.
• The activation threshold test $\sum_{i=1}^n \bar w_i \bar x_i \geq \theta$ is converted to $\{0,1\}$0, where $\{0,1\}$1 is the set of active input literals and $\{0,1\}$2 is a scaled threshold.
• The capacity $\{0,1\}$3 is the indicator of subsets of $\{0,1\}$4 of size at least $\{0,1\}$5, and the Sugeno integral $\{0,1\}$6 encodes the activation logic.

This construction is compositional: each hidden layer can be encoded and processed in the same way, and the output of one layer naturally feeds as "polarized" binary input to the next.

Last-Layer Aggregation and Real-Valued Output

The last layer, which generally produces real-valued scores (logits), is represented similarly. The capacity in this setting is rescaled to measure the proportion of matching inputs, and the output is given as an affine function of the Sugeno integral of the matching literal pattern. This preserves the original neural computation exactly on the encoded binary input.

Rule-Based and Explanation Semantics

A direct consequence is that every BNN neuron corresponds to a set of minimal activation rules determined by the focal sets of the capacity. The Sugeno integral representation thus exposes the logical structure of activations, facilitates counterfactual reasoning (minimal changes to cross thresholds), and provides an explicit mapping to the collection of if-then rules realized by the neuron or layer.

Theoretical and Practical Implications

Explainability and Verification

By recasting hidden neuron activations in terms of focal sets and capacity evaluation, the approach provides a transparent path toward local explanations, sensitivity analysis, and formal verification for BNNs. It gives an explicit characterization of all minimal patterns that suffice or are necessary for activation, in contrast to the opaque nature of raw weighted sums. This has immediate consequences in settings requiring post-hoc semantic interpretation of neural decisions, such as safety-critical applications, and connects symbolic and subsymbolic views in a rigorous manner.

Extension to Richer Capacities and Multivalued Encodings

The mapping leverages the fact that BNNs use binary weights and activations, but the Sugeno framework is substantially more general. The authors emphasize that more general (non-additive, interactive) capacities can represent richer dependencies among inputs, and thus the framework naturally extends to quantized networks where activations and/or weights assume more than two discrete values [e.g. 8-bit quantization as in Jacob et al.]. The correspondence also opens the door to representing network computations on finite ordinal scales, allowing principled interpolation between symbolic rule-based logic and standard deep learning.

Connection to Neurosymbolic Systems and Future Work

This capacity-centric viewpoint aligns with the growing body of research in neurosymbolic AI, which seeks to embed logical reasoning capabilities within neural architectures. The paper references recent work defining neurosymbolic inference as a form of "integral" over logical and probabilistic functions and positions the Sugeno integral as a candidate formalism for future developments in this area. The explicit link to rule induction and focal set semantics also provides new approaches for learning interpretable fuzzy measures or capacities from data, suggesting an intersection with learning theory for explainable and robust AI.

Conclusion

This work delivers a precise and systematic translation from BNN inference to the Sugeno integral framework, revealing the minimal combinatorial structure underlying each neuron and extending to the network’s output layer. The immediate theoretical implications cover explainability, verification, and logical rule extraction; practically, they set the stage for methodical integration of qualitative reasoning within neural architectures and for further research in quantized and neurosymbolic systems. The Sugeno integral’s role as both an aggregator and a transparent logic operator situates it as an effective tool for advancing both the interpretability and the expressivity of contemporary AI models.