Kolmogorov-Arnold Networks for Personal Context Recognition on ExtraSensory

Hoang-Thang Ta

cs.LGOct 4, 2026

Kolmogorov-Arnold Networks for Personal Context Recognition on ExtraSensory

Authors: Hoang-Thang Ta

Organizations: University of Information Technology, Ho Chi Minh City, Vietnam · Vietnam National University Ho Chi Minh City, Ho Chi Minh City, Vietnam

Abstract

Kolmogorov--Arnold Networks (KANs) have attracted increasing attention in recent years, with applications across a wide range of AI tasks. In this paper, we evaluate several KAN variants on the ExtraSensory dataset for personal context recognition and compare them with a multilayer perceptron (MLP) and TabM. We conduct the main experiments using five user folds and three random seeds per fold and report the average Macro-F1, Micro-F1, and training time. We also perform shallow ablation studies on grid size, the number of grids, and data normalization to examine their effects on KAN performance. The results show that all evaluated KAN variants significantly outperform MLP in terms of Macro-F1 and Micro-F1 and achieve performance comparable to TabM. However, KAN variants generally require more training time, while TabM provides a more favorable balance between predictive performance and training efficiency. These results suggest that KANs are promising for personal context recognition, while their computational efficiency remains an important challenge. Our source code is publicly available at: https://github.com/hoangthangta/ExtraSensory-KANs.

Figures & tables

Figure 1: Left: The architecture of KAN(2,3,1). Right: Calculation of

\phi_{1,1,1}

using control points and B-spline basis functions, where

G

and

k

denote the grid size and spline order, respectively, and

n=G+k=3+3=6

is the number of B-spline basis functions. The figure is reused from Ta [23] .

Variant	Basis Function
EfficientKAN [ 4 ]	$\phi(x)=w_{b}b(x)+w_{s}B(x)$ , where $B(x)=\sum_{i}c_{i}B_{i}(x)$ , with $c_{i}$ denoting spline coefficients and $B_{i}(x)$ denoting B-spline basis functions.
FastKAN [ 14 ]	$\phi(x)=w_{b}b(x)+w_{s}\mathit{RBF}(x)$ , where $\mathit{RBF}(x)=\exp\!\left(-\frac{\\|x-d\\|^{2}}{2h^{2}}\right)$ ; $d$ is the center and $h$ is the function width or spread.
BSRBF-KAN [ 22 ]	$\phi(x)=w_{b}b(x)+w_{s}\big(\mathit{B}(x)+\mathit{RBF}(x)\big)$ , where $\mathit{B}(x)$ and $\mathit{RBF}(x)$ are defined above.

Table 1: Summary of KAN variants and their basis functions.

Model	Params	Network structure
BSRBF-KAN	72,544	$225$ – $29$ – $51$ ; grid size $=5$ , spline order $=3$ , LayerNorm, SiLU.
EfficientKAN	71,760	$225$ – $26$ – $51$ ; grid size $=5$ , spline order $=3$ , SiLU.
FastKAN	72,624	$225$ – $29$ – $51$ ; grid size $=8$ , LayerNorm, SiLU.
MLP	71,925	$225$ – $256$ – $51$ ; LayerNorm and SiLU.
TabM	71,002	225 numerical features; 1 block, $k=8$ .
Params = number of trainable parameters.

Table 2: Network structures and numbers of trainable parameters for the models on the ExtraSensory dataset.

Model	Macro-F1	Micro-F1	Train Time (s)
BSRBF-KAN	$0.3106\pm 0.0095$	$\mathbf{0.5572\pm 0.0123}$	665.47
EfficientKAN	$\mathbf{0.3112\pm 0.0072}$	$0.5556\pm 0.0179$	540.80
FastKAN	$0.3091\pm 0.0111$	$0.5527\pm 0.0143$	463.80
MLP	$0.2807\pm 0.0099$	$0.4957\pm 0.0184$	$\mathbf{372.86}$
TabM	$0.3107\pm 0.0089$	$0.5562\pm 0.0087$	381.30

Table 3: Multi-label classification results on ExtraSensory. Each model is evaluated over 15 runs (5 folds and 3 random seeds). Results are reported as mean

\pm

standard deviation, and training time is reported as the average over those runs.

Metric	Comparison	$p$	$p_{\mathrm{adj}}$	Sig.
Macro-F1	BSRBF-KAN vs. EfficientKAN	0.6387	1.0000	No
	BSRBF-KAN vs. FastKAN	0.5995	1.0000	No
	BSRBF-KAN vs. MLP	0.0001	0.0012	Yes
	BSRBF-KAN vs. TabM	0.8469	1.0000	No
	EfficientKAN vs. FastKAN	0.5995	1.0000	No
	EfficientKAN vs. MLP	0.0001	0.0012	Yes

Table 4: Pairwise Wilcoxon signed-rank tests for multi-label classification. The tests are conducted over 15 matched fold–seed runs, and

p

-values are adjusted using Holm correction.

Figure 2: Ablation study of grid size and the number of grids ( num_grids ) in BSRBF-KAN, EfficientKAN, and FastKAN. Each model was evaluated on fold 0 using three random seeds (0, 1, and 2). The average Macro-F1 across the three seeds was normalized to the range $[0,1]$ for visualization in the heatmap, where 0 and 1 denote the lowest and highest results, respectively. For BSRBF-KAN and EfficientKAN, spline_order = 3 and grid_size $\in\{2,3,5,8\}$ . For FastKAN, num_grids $\in\{5,6,8,11\}$ , corresponding to grid_size $\in\{2,3,5,8\}$ under num_grids = grid_size + spline_order .

Model	Normalization	Macro-F1
BSRBF-KAN	BatchNorm	$\mathbf{0.3214\pm 0.0058}$
	LayerNorm	$0.3168\pm 0.0177$
	No Norm	$0.3153\pm 0.0113$
FastKAN	LayerNorm	$\mathbf{0.3190\pm 0.0013}$
	No Norm	$0.3089\pm 0.0051$

Table 5: Macro-F1 results for various normalization strategies on ExtraSensory. Results are averaged over three seeds and reported as mean

\pm

standard deviation.

Explore similar work

Kolmogorov--Arnold Networks for Small Language Models
Jul 17, 2026Felippe Alves, Renato VicenteKolmogorov-Arnold Networks Multilayer Perceptrons
TELUS Digital Research Hub – CIAAM Institute of Mathematics, Statistics and Computer Science University of São Paulo
Evaluating the Effectiveness of SechKAN on 1D Data
Sep 22, 2026Hoang-Thang TaKolmogorov-Arnold Networks Human Activity Recognition
University of Information Technology, Ho Chi Minh City, Vietnam · Vietnam National University Ho Chi Minh City, Ho Chi Minh City, Vietnam
SparseKAN: Compressing Kolmogorov--Arnold Networks Across Basis Functions, Neurons, and Bits
Aug 1, 2026Kazi Ahmed Asif Fuad, Lizhong ChenKolmogorov-Arnold Networks Compressed Model
Department of EECS Oregon State University Corvallis, OR 97331

Jul 17, 2026, cs.LGJ/K move · Enter open · S save

Kolmogorov--Arnold Networks for Small Language Models

Felippe Alves, Renato Vicente

TELUS Digital Research Hub – CIAAM Institute of Mathematics, Statistics and Computer Science University of São Paulo

Kolmogorov-Arnold Networks Multilayer Perceptrons

Open paper AskPDF arXiv

Kolmogorov--Arnold Networks (KANs) replace fixed node activations with learned one-dimensional edge functions, offering an explicit interface for interpretation and a possible alternative to transformer feed-forward networks. We test these claims separately. In a six-layer, 10M-parameter B-spline KAN, we reconstruct all 884,736 feed-forward edges: 87.8% exceed (NLS>0.1) and 0.4% are inactive. Pruning the lowest-activity 20--25% causes negligible loss increase, although structured MLP neuron pruning tolerates comparable sparsity. The audit replicates on BabyLM, but grid-size sweeps show that near-total fPCA compression and high closed-form-fit coverage are properties of the low-capacity grid-2 basis, not universal KAN behavior. For replacement, we evaluate MLP, SwiGLU, grouped Chebyshev, and rational GR-KAN networks on BabyLM. The KAN-family and gated variants improve validation loss over the GELU MLP, but this ordering does not transfer to standardized benchmarks: across ten seeds and 59,875 BLiMP pairs, accuracies span 62.4--63.1%, EWoK remains at chance, and a (+0.7)-point GR-KAN effect on BLiMP reverses on the supplement. Larger tests are also cautionary: parameter-matched MLPEdge underperforms the MLP on Wikitext-103, and 286M-parameter GR-KAN remains below a SwiGLU ClimbMix baseline after stabilization. Thus, small-basis KANs provide a practical, corpus-transferable interface for auditing learned scalar transformations, but the tested replacements show no consistent benchmark, quality, or latency advantage over strong MLP baselines.

Abstract

Explore similar work

Kolmogorov--Arnold Networks for Small Language Models

Evaluating the Effectiveness of SechKAN on 1D Data

SparseKAN: Compressing Kolmogorov--Arnold Networks Across Basis Functions, Neurons, and Bits