The work, published in the journal Complex & Intelligent Systems, comes from an international team spanning SRM Institute of Science and Technology in India, the National Institute of Technology Rourkela, China University of Mining and Technology, Innopolis University in Russia, and St. Petersburg Electrotechnical University “LETI.” The team — M. R. Indresh, Soumyajit Gayen, Dmitrii Minenkov, Dmitrii Kaplun and Ram Sarkar — describes FAED, a Fuzzy Attention-aided Encoder-Decoder architecture, which swaps out the rigid binary logic of standard attention mechanisms for a soft, probabilistic notion of relevance inspired by fuzzy set theory. The results are striking not only for their accuracy but for the architecture’s remarkable frugality: with just 2.4 million parameters and roughly 4 GFLOPs of computation, FAED posts Dice scores that put it at the top tier of contemporary segmentation models while running at inference speeds measured in milliseconds.

The clinical stakes of this problem are easy to underestimate. Early-stage melanoma is highly curable, but the first line of defense is visual inspection of pigmented lesions, typically through dermoscopy — the imaging of skin through a magnifying device that reveals subsurface structures. Automated segmentation of dermoscopy images, the task of drawing an accurate boundary around a lesion, underpins every downstream measurement clinicians and computer-aided diagnosis systems rely on, including the asymmetry, border irregularity and color variation criteria used in melanoma risk scoring. Yet the task is plagued by low contrast between lesion and healthy skin, hair occlusions, specular reflections, and most fundamentally, ambiguous boundaries where the lesion fades gradually into its surroundings.

For years, the dominant tool for this job has been U-Net, a convolutional encoder-decoder architecture in which a contracting path extracts increasingly abstract features and an expanding path reconstructs a pixel-level prediction. The critical link between the two halves is a set of skip connections that pass fine-grained spatial detail from early encoder layers directly to the decoder. Most modern variants bolt attention modules onto these skip connections: the network learns to “gate” which features to pass through. But those gates are typically binary — a feature channel or spatial position is either kept or discarded. The FAED authors argue that this all-or-nothing logic is fundamentally mismatched to the nature of skin lesions, where the transition between sick and healthy tissue is gradual, not sharp. A binary gate discards exactly the soft, intermediate evidence that defines an ambiguous boundary.

FAED’s central innovation is its Boundary-conditioned Soft Fuzzy Attention (BSFA) module, which replaces standard skip connections altogether. Rather than multiplying features by a learned 0-or-1 mask, BSFA evaluates feature relevance using learnable Gaussian membership functions — mathematical constructs from fuzzy logic that assign each feature a continuous degree of membership, modeled as a probability-like value between zero and one. In practice, this means the network can express that a feature is “somewhat relevant” or “mostly relevant,” preserving graded boundary information that binary attention would crush. The Gaussian membership functions are themselves learnable parameters, so the network discovers its own notions of partial relevance during training rather than having them imposed by a fixed rule.

The architecture adds two further refinements that the authors show are individually and jointly important. The first is an Adaptive Fuzzy Mixture-based aggregation scheme. Features extracted at different depths of the network vary enormously in scale and semantics — shallow layers carry edge textures, deep layers carry abstract shape information — and fusing them well is a persistent headache in segmentation design. The fuzzy mixture approach treats each feature source as contributing to a soft ensemble, weighting its contribution according to a learned similarity-based membership rather than simple concatenation. The second refinement is an explicit Boundary Cue, a signal fed into the attention mechanism that modulates its focus along lesion perimeters. Where the fuzzy membership decides “how relevant” a feature is, the boundary cue tells the attention “where to look,” sharpening the model’s sensitivity precisely at the lesion border where errors are most costly.

The authors validated FAED on the four most widely used benchmarks in the field: the ISIC2016, ISIC2017 and ISIC2018 dermoscopy datasets from the International Skin Imaging Collaboration, and the smaller PH² dataset of melanocytic lesion images. The segmentation quality was measured with the Dice score, a standard metric that quantifies the overlap between the predicted lesion mask and the ground truth, where a score of 1.0 means perfect agreement. FAED achieved a Dice score of 0.9140 on ISIC2016, 0.9135 on PH², 0.8781 on ISIC2018, and 0.8615 on ISIC2017 — competitive-to-leading figures given the architecture’s size. Notably, the ISIC2016 and PH² results hover around the 0.91 mark, a level of overlap that corresponds to clinically meaningful boundary fidelity.

Just as important as the headline numbers is the efficiency profile, which the team documented with careful empirical measurements on an NVIDIA Tesla T4 GPU. FAED performs inference in 10.05 milliseconds per image at batch size 1, and 6.76 milliseconds per image when batched at 8 — throughput fast enough for real-time clinical workflows. Peak GPU memory during inference is similarly modest: 505 MB at batch size 1 and 948 MB at batch size 8. For context, many state-of-the-art segmentation models rely on heavyweight transformer backbones or large convolutional stacks with parameter counts in the tens of millions, demanding memory and compute budgets that make deployment on hospital hardware, edge devices or low-resource settings difficult. FAED’s 2.4 million parameters and roughly 4 GFLOPs place it in a different class entirely, suggesting that careful architectural design — rather than brute-force scale — can carry segmentation performance a long way.

To verify that each component of the design actually earns its place, the researchers conducted ablation studies, the standard experimental practice of removing parts of a system one at a time and measuring the drop in performance. These studies confirmed that both the prototype-based fuzzy aggregation and the boundary-conditioned modulation of attention contribute measurably to the observed improvements. In other words, the gains are not an artifact of added capacity or incidental tuning: the soft membership modeling and the explicit boundary guidance are doing real, distinguishable work. That finding matters for the broader field, because it offers evidence that how features are fused — treating fusion as a soft similarity-based membership problem — can be as consequential as how features are extracted.

The philosophical shift at the heart of FAED is worth dwelling on. Classical computer vision and early deep learning systems were built on crisp logic: a pixel belongs to a class, a feature passes a gate, a decision is yes or no. Fuzzy logic, introduced decades ago as a formal way of reasoning with degrees of truth, has long been touted as a natural fit for medical imaging, where human experts themselves reason in gradients — “this border looks slightly irregular,” “this region is probably part of the lesion.” What has changed recently is that learnable fuzzy components, such as Gaussian membership functions optimized end-to-end by gradient descent, can now be embedded inside deep networks so that the fuzzy rules themselves are discovered from data. FAED is a concrete demonstration that this marriage of classical soft-computing theory and modern deep learning can outperform hard-gated alternatives on a real clinical task, without any increase in architectural complexity.

The implications for melanoma screening are potentially significant, particularly for parts of the world where dermatologists are scarce and mobile screening programs depend on lightweight, fast and reliable algorithms. A model that runs in a few milliseconds on an entry-level GPU, fits comfortably in under a gigabyte of memory, and still achieves over 91 percent overlap with expert-annotated boundaries on benchmark datasets is precisely the kind of tool that can be embedded into telemedicine pipelines or portable dermoscope accessories. The authors caution, as all careful researchers do, that benchmark performance is a step toward clinical deployment, not the deployment itself — prospective validation on diverse skin tones, imaging devices and real-world lesion appearances remains an essential next stage for any segmentation technology destined for the clinic.

The article was published open access under a Creative Commons license, making the full technical description freely available to researchers and clinicians worldwide. The study was supported by the Ministry of Economic Development of the Russian Federation. As peer-reviewed, citable research made available early for faster dissemination, it joins a growing body of work arguing that the future of medical AI lies not only in ever-larger models, but in smarter ones — systems that, like the physicians they assist, know how to say “maybe.” FAED’s fuzzy attention may be an early but compelling example of that principle turned into working code.

At the heart of the development is a question that has occupied mathematicians and computer scientists for decades: how can a machine approximate a complicated function of many variables? A conventional multilayer perceptron, or MLP, processes information through layers of artificial neurons. Each neuron applies a weighted sum to its inputs and then passes the result through a nonlinear activation function such as a rectified linear unit. The network learns the weights connecting neurons, gradually bending a high-dimensional input-output relationship into a useful approximation. KANs take a different route. Inspired by the Kolmogorov-Arnold representation theorem, they place learnable one-dimensional functions on connections between nodes rather than relying primarily on fixed activation functions inside neurons. In principle, this lets the model build a complex multivariable function by combining simpler functions of individual coordinates. The approach has generated intense interest because the learned functions can sometimes be inspected directly, offering a possible advantage in interpretability over opaque deep networks.

The mathematical theorem behind KANs does not automatically guarantee that every practical implementation will work well. The original Kolmogorov-Arnold representation concerns continuous functions and provides an existence result, but translating that insight into a stable, efficient learning algorithm is a separate engineering and analytical challenge. Real optimization problems often involve functions with sharp bends, kinks, regime changes or other forms of irregularity. Such behavior can arise when a system switches between operating constraints, when a physical process changes abruptly, or when a small variation in an input produces a disproportionately large change in the output. Smooth spline-based KANs can be highly effective when the target function is itself smooth, but their performance may degrade when the function contains irregular structure. Warin’s P1-KAN is designed around this difficulty, using piecewise-linear components intended to represent abrupt changes without requiring the network to force them into an overly smooth shape.

The “P1” designation refers to first-order, piecewise-polynomial behavior. Rather than representing each learnable connection with a globally smooth curve, the architecture constructs functions from local linear segments. A piecewise-linear function changes slope at selected breakpoints, allowing it to approximate a sudden transition while retaining a comparatively simple computational form. This design can also make optimization more manageable: the model does not need to adjust a large collection of highly flexible smooth basis functions to reproduce a sharp feature. P1-KAN therefore occupies a different point in the trade-off between flexibility, accuracy and computational cost. The study presents universal approximation theorems for several versions of the architecture, meaning that, under specified conditions and with sufficient model capacity, the networks can approximate broad classes of target functions to arbitrary precision. The authors also derive error estimates for cases in which the underlying Kolmogorov-Arnold representation functions possess regularity, providing a theoretical framework for understanding when the model should perform well.

To test the practical consequences of the theory, the researchers first examined simple regression problems in which the networks had to learn known mathematical relationships. These controlled experiments make it possible to separate the architecture’s behavior from the complications of a large industrial data set. According to the study, P1-KAN achieved higher accuracy than MLPs and reached useful solutions more quickly during training. In machine learning, convergence speed measures how rapidly an optimization procedure reduces its error or objective function. Faster convergence can lower computational expense and can be crucial when a model must be retrained repeatedly as conditions change. The results also showed that P1-KAN was especially effective when the target functions were irregular. For smooth functions, its accuracy was similar to that of the original spline-based KAN, suggesting that the piecewise-linear design does not sacrifice performance simply because it is built to handle rougher mathematical terrain. The comparisons included several other KAN variants, reflecting the rapidly expanding ecosystem of architectures based on the same broad idea.

The decisive demonstration involved the optimization of a hydraulic valley in France, an industrial problem connected to the operation of water reservoirs and hydropower infrastructure. Reservoir optimization requires decisions about how water should be stored and released over time. Operators must balance competing objectives, such as electricity production, water availability, downstream constraints and the uncertain arrival of future inflows. A release that maximizes power generation today may reduce flexibility tomorrow; retaining water may preserve future options but miss a valuable opportunity in the present. Mathematically, the problem can be framed as a sequential control task in which the optimal action depends on the current state of the hydraulic system and on uncertain future conditions. The value function—the estimated long-term benefit associated with a particular state—can become highly nonlinear and irregular, especially when operational limits or discrete decisions are involved. That makes it an exacting environment for a function-approximating neural network.

Practitioners have traditionally approached such problems with dynamic programming, a method formalized in influential work on sequential decision-making and stochastic control. Dynamic programming breaks a complex multistage problem into linked subproblems and uses a recursive relationship to calculate the value of decisions over time. In a simple setting, the method can be extraordinarily powerful. But its computational demands grow rapidly as the number of state variables increases, a difficulty commonly described as the curse of dimensionality. Fine-grained representations of reservoir levels, inflows, market conditions and other variables can require enormous memory and processing time. Neural networks offer a way to approximate the value function or policy without explicitly enumerating every possible state. They can generalize from sampled scenarios, potentially making optimization feasible in settings where a grid-based dynamic-programming calculation becomes unwieldy. Yet that advantage depends on the network learning the system’s irregularities rather than smoothing them away or converging to a misleading solution.

In the hydraulic valley experiment, P1-KAN optimized the system more effectively than the other KAN networks tested and also surpassed the classical deterministic tools used by practitioners, the study reports. The result does not mean that a neural network has replaced hydraulic expertise or eliminated uncertainty from reservoir management. Instead, it indicates that the architecture’s particular way of representing nonlinear functions may be valuable for a difficult class of industrial optimization problems. A model can perform well in this setting because it captures the structure of the objective function and constraints with fewer approximation errors, because it trains more efficiently, or because its local piecewise representation adapts better to changes in operating regimes. The article’s findings support the idea that architecture matters: simply applying a larger or deeper network is not always the best answer. Matching the mathematical properties of a model to the structure of the problem may deliver larger gains than adding layers or parameters.

The implications extend well beyond water management. Kolmogorov-Arnold networks are being explored for mechanics, medical-image analysis, time-series forecasting, survival analysis, physics-informed neural networks and other applications in which interpretability or nonlinear approximation is important. P1-KAN could be relevant wherever a system contains thresholds, discontinuities or sharply changing responses, from energy dispatch and infrastructure planning to stochastic control and engineering design. At the same time, the study highlights why enthusiasm should be paired with careful validation. The experiments described in the article establish theoretical approximation properties and report strong results on selected regression and hydraulic optimization tasks, but they do not prove that P1-KAN will outperform every architecture on every data set. Other research has reported limitations for KANs on noisy functions, and comparisons between neural-network families can depend heavily on training procedures, parameter counts, data quality and hardware implementation. The authors state that data are available from them on request, which could allow independent researchers to examine the benchmarks and reproduce the findings.

The broader significance of P1-KAN is that it brings a more problem-aware philosophy to the current neural-network race. Artificial intelligence has often advanced by scaling: more data, more parameters and more computing power. But many scientific and industrial systems are governed by equations, constraints and abrupt physical or economic transitions that generic architectures may represent inefficiently. A network built from adaptable one-dimensional functions can expose a different set of mathematical building blocks, while a piecewise-linear version can target irregularity directly. Warin’s results suggest that this combination can make a measurable difference, both in clean mathematical tests and in a real optimization problem with practical consequences. If future studies confirm the gains across additional reservoirs, uncertainty models and high-dimensional control tasks, P1-KAN may become part of a new generation of scientific machine-learning tools—systems designed not merely to fit data, but to respect the shape of the problems they are asked to solve.