⚡ electrical engineering

Distance-Aware Error for Spline Networks: A Bottom-Up Approach to Uncertainty

This contribution presents a deterministic, bottom-up approach to deriving distance-dependent error bounds in spline neural networks, including Kolmogorov-Arnold networks, by analyzing individual neuron errors and their propagation through compositions, thereby demonstrating superior speed and reliability compared to probabilistic baselines in applications such as shape estimation and safe navigation.

Original authors: Masoud Ataei, Mohammad Javad Khojasteh, Vikas Dhiman

Published 2026-05-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Masoud Ataei, Mohammad Javad Khojasteh, Vikas Dhiman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to navigate a dark forest. You give the robot a map based on a few trees it has seen before. When the robot is near those trees, it feels safe. But what happens when it wanders into a part of the forest it has never seen before? An intelligent robot should not simply guess; it should say: "I am not sure here; be careful."

This article introduces a new tool called DAREK (Distance-Aware Uncertainty for Kolmogorov Networks), which helps these "Spline Neural Networks" (a type of AI that uses smooth, flexible curves for learning) know precisely how uncertain they should be.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "Overconfident Fool"

Most AI models are like a student who has memorized the answers to a practice test. If you ask them a question from that test, they get it right. But if you ask them a question from a completely different topic, they might still give an answer, yet they do not know it is wrong. They lack a "safety cage."

In safety-critical tasks (such as driving a car or controlling a robot), this is dangerous. We need a system that knows: "The further I am from the data I was trained on, the less I should trust my own answer."

2. The Solution: Building a Safety Net from the Bottom Up

The authors propose a "bottom-up" approach. Instead of estimating the uncertainty of the entire network all at once, they calculate the uncertainty of each individual tiny part (neuron) and then stack these together.

Imagine the AI network as a relay race team:

The Neurons (Runners): Each neuron takes a piece of information, smooths it, and passes it to the next runner.
The Error (The Stumble): Sometimes a runner stumbles a little.
The DAREK Method: Instead of waiting to see if the entire team finishes the race incorrectly, DAREK calculates exactly how much each runner could stumble, based on how far they are from the training run. It then adds up all these potential stumbles to tell you the maximum possible error for the entire team.

3. How it Measures "Distance"

The article uses a clever mathematical trick with knots.

The Analogy: Imagine a flexible wooden strip (a spline) held in place at specific points by nails (knots). The strip is smooth between the nails.
The Rule: If you stand right next to a nail, the strip is very predictable. If you stand exactly between two nails, you have to guess how the wood bends. If you are far away from any nail, your prediction is very risky.
DAREK's Task: It measures how far your current question is from the nearest "nail" (training data). The further away you are, the wider the "safety cage" (error bound) becomes. This ensures that the AI admits uncertainty when it is in unfamiliar terrain.

4. Why it is Better Than Other Methods

The article compares DAREK with two other popular methods for estimating uncertainty:

Monte Carlo (The "Rolling Dice" Method): This tries to run the AI thousands of times with slight random changes to see how much the answer varies. It is accurate but slow, like asking 20 different people to solve a math problem and averaging their answers.
Gaussian Processes (The "Statistical Map" Method): This uses complex statistics to draw an uncertainty map. It is good but becomes heavy and slow when you add more data, like trying to carry a huge, detailed atlas everywhere you go.

DAREK's Advantage: It is fast and deterministic. It does not need to roll dice or carry a heavy atlas. It uses a direct mathematical formula to calculate the "worst-case scenario" immediately. The article shows that it is faster than both the "rolling dice" method and the "statistical map" method.

5. Real-World Tests

The authors tested DAREK in three specific scenarios mentioned in the article:

Shape Guessing: Trying to guess the shape of an object from a few scattered laser points. DAREK knew when it was guessing too far from the points.
Face Recognition: Finding a face in an image with many features. DAREK remained reliable even when the image was complex.
Safe Navigation: A simulation where robots had to move without colliding. When the robots used DAREK, they successfully avoided collisions in tricky situations where other methods failed because they were too overconfident in their wrong guesses.

Summary

DAREK is a new "safety calculator" for a certain type of AI. It works by measuring how far the AI is from the data it learned from.

Close to the data? The AI is confident.
Far from the data? The AI raises a big red flag and says: "I do not know this well; be careful."

It does this quickly, without needing to run thousands of simulations, making it perfect for robots and systems where safety is the top priority. The authors have even provided a free library so other engineers can build this "safety cage" into their own AI models.

Technical Summary: Distance-Aware Error for Spline Networks (DAREK)

Problem Statement
The work addresses the challenge of uncertainty quantification in Spline Neural Networks (SNNs), particularly Kolmogorov-Arnold Networks (KANs). While probabilistic methods (e.g., Monte Carlo, Bayesian Neural Networks) and Gaussian Processes (GPs) exist, they often suffer from high computational costs, dependence on sampling, or lack of interpretability. Furthermore, existing worst-case error bounds for KANs are often too loose, not distance-aware (i.e., they do not increase monotonically with distance from training data), or defined only up to an unknown constant. The authors aim to develop a deterministic, distance-aware, and tight error bound for SNNs that is computationally efficient and suitable for safety-critical applications where probabilistic assumptions may not hold.

Methodology
The proposed method, DAREK (Distance-Aware uncertainty for Kolmogorov network), employs a bottom-up analytical approach to derive worst-case error bounds. The methodology proceeds in three main phases:

Neuron-level Error Bounds:
- The analysis begins by treating each neuron in the network as a spline function.
- The authors utilize error bounds for Newton polynomials (Theorem 1) to establish a tight bound for piecewise polynomial approximations, assuming the underlying function is $(k+1)$ -times Lipschitz continuous. This bound is inherently distance-aware and scales with the geometric mean of the distances to the $k+1$ nearest knots.
- To handle real-world scenarios where the spline approximation does not pass exactly through the knots (due to noise or optimization constraints), the authors extend the bound via Lemma 1. This integrates the error at the knots, approximated via a piecewise linear fit (Error Bound with Linear Fit, EBL), to avoid the Runge phenomenon and numerical instabilities.
Layer-wise and Network-wide Propagation:
- The error bounds are extended to multi-layer networks by analyzing the composition of spline functions.
- Theorem 2 and Theorem 3 provide error bounds for two-layer and $L$ -layer KANs, respectively. The total error comprises the interpolation error of the current layer and the propagated error from previous layers, scaled by the Lipschitz constants of subsequent layers.
- The framework handles residual connections (e.g., SiLU functions added to splines in KANs) by treating the residual as an error-free component and attributing the total error to the spline part (Theorem 4).
Assignment of Constants and Errors:
- A critical challenge lies in distributing the global Lipschitz constant and the total approximation error across individual neurons and layers. The work proposes and evaluates several strategies:
  - Lipschitz Distribution: Methods include optimization-based, logarithmic, linear, equal, and worst-case distributions. Empirical results suggest that the "Equal Distribution" offers a favorable trade-off between computational cost and bound tightness.
  - Error Distribution at Knots: Since layer-wise errors are unobservable, the total error at the knots is distributed using methods such as "Last Layer Error," "Equal Distribution," or "Shapley Values" (approximated via Monte Carlo sampling).

Main Contributions

Analytical Worst-Case Bounds: The work establishes the first analytical, distance-aware error bounds for multi-layer spline neural networks under assumptions of higher-order Lipschitz continuity.
Bottom-up Composition: It introduces a novel framework for composing error bounds from individual neurons to the entire network, explicitly handling error propagation and knot errors.
Extension to Smooth Functions: The method extends error bounds to layers containing additive smooth functions (e.g., SiLU) alongside splines.
Distance-Awareness Metric: A metric (Sampled Distance-Awareness, SDA) was developed to quantitatively evaluate how well an uncertainty estimator increases with distance from training data.
Efficiency and Interpretability: The approach provides a library for integrating these bounds into existing KAN models and offers a computationally efficient alternative to GPs and Monte Carlo ensembles.

Experimental Results
The authors evaluated DAREK on several tasks:

Function Approximation: On benchmark functions (e.g., Cosine, Two-Moon dataset), DAREK successfully enclosed the true error. The "Equal Distribution" strategy for Lipschitz constants yielded a low violation rate with minimal computational overhead.
Object Shape Estimation: In a 2D laser scanning task, DAREK estimated the Signed Distance Function (SDF) of an object, demonstrating its ability to generalize to real-world mapping problems.
High-Dimensional Face Recognition: On the CelebA dataset, DAREK was compared with GPs and ensembles. While GPs achieved perfect SDA due to their kernel formulation, DAREK's SDA improved significantly with dimension (reaching ~95% at 100 dimensions) and outperformed ensemble methods, which exhibited inconsistent patterns.
Safe Multi-Agent Navigation: In a simulated environment with bounded disturbances, DAREK was integrated into a robust control framework (MPC + CBF). DAREK-based controllers achieved success rates of up to 95%, outperformed a primal controller, and matched or exceeded GP-based controllers in safety-critical navigation tasks, while being more computationally efficient.

Significance and Claims
The work claims that DAREK offers a unique combination of interpretability, computational efficiency, and theoretical guarantees. By localizing errors to specific knots and ensuring that bounds are distance-aware, the method provides a "safety cage" for SNNs, making them suitable for safety-critical domains such as robotics and autonomous control. The authors emphasize that their bounds are deterministic and require neither probabilistic assumptions nor large sample sizes for inference, thereby addressing a major limitation of Bayesian and GP approaches. However, the authors modestly acknowledge limitations, such as dependence on training data for knot selection (which may limit generalization if the sample is not representative) and the potential for excessive conservatism when aggregating bounds across high dimensions. The authors note that the proposed Lipschitz and error distribution methods are heuristics and are not guaranteed to perfectly align with true system properties.