How We Used Kolmogorov-Arnold Networks to Classify Crayfish — And Got Published in Nature

Earlier this year, our paper was published in Scientific Reports (Nature). It was my first academic publication, and it came from a research area I didn’t expect to end up in: crayfish biology.

Here’s what we did, why it matters, and what I learned along the way.

The problem

Determining the sex of crayfish is important for population management and ecological research. Traditional methods rely on manual inspection of morphological features — which is time-consuming, requires expertise, and isn’t always accurate, especially with younger specimens.

The question was simple: can machine learning do this better?

The approach

We used two types of datasets — tabular measurements (weight, size, morphological characteristics) and image data (photographs of each specimen). The goal was binary classification: male or female.

What made this project interesting was the model choice. Instead of defaulting to standard architectures, we explored Kolmogorov-Arnold Networks (KAN) — a relatively new approach that replaces fixed activation functions with learnable ones on the edges of the network.

What are KANs?

Traditional neural networks use fixed activation functions (ReLU, sigmoid) at each node. KANs flip this — they place learnable univariate functions on the edges instead. This is based on the Kolmogorov-Arnold representation theorem, which states that any multivariate continuous function can be decomposed into compositions of univariate functions.

In practice, this means KANs can capture complex nonlinear patterns with fewer parameters and better interpretability than standard MLPs.

The full pipeline

We didn’t just use KANs in isolation. The pipeline included:

Stacked Autoencoders for dimensionality reduction and feature extraction
KAN as the primary classifier
ANN, KNN, SVM, Random Forest, and Logistic Regression as baselines for comparison
Cross-validation for robust evaluation

The results

The KAN model achieved 95–100% accuracy depending on the dataset and configuration. It consistently outperformed or matched the traditional classifiers, while offering better insight into which features mattered most for classification.

The stacked autoencoder preprocessing was critical — it compressed the feature space while preserving the signal, giving the downstream classifiers cleaner inputs.

What I learned

Research is slow, engineering is fast. In software engineering, you ship features in days. In research, a single experiment can take weeks to design, run, and analyze properly. The pace is completely different, and both have value.

Model selection matters less than data preparation. We spent more time on data collection, cleaning, and feature engineering than on model architecture. The autoencoders did the heavy lifting — the classifier on top was almost secondary.

KANs are worth watching. They’re still early, but the combination of interpretability and performance is promising. For tabular data especially, they offer something that black-box deep learning models don’t: a window into what the model actually learned.

Publishing is a team effort. This paper had nine authors across multiple institutions. Coordinating experiments, writing, and revisions across that many people taught me more about collaboration than any sprint planning session.

The paper

If you want to read the full paper:

Y. Atilkan, B. Kirik, E. T. Acikbas, F. Ekinci, K. Acici, T. Asuroglu, R. Benzer, M. S. Guzel, and S. Benzer. “Enhancing crayfish sex identification with Kolmogorov–Arnold networks and stacked autoencoders.” Scientific Reports, Nature, 2025.

Read the full paper →

This was my first publication, and it confirmed something I’d been feeling for a while: the intersection of software engineering and AI research is where I want to be. My M.S. at Ankara University is pushing me further in that direction, and I’m currently working on a second paper applying similar techniques to forensic anthropology.