Revisiting Multilayer Perceptrons in the Era of Kolmogorov Arnold Networks: An Empirical Comparison on Image Classification

Multilayer perceptrons (MLPs) remain a widely used architecture in deep learning, whereas Kolmogorov–Arnold networks (KANs) have emerged as an alternative that represents nonlinear transformations through learnable univariate functions, commonly parameterized using splines. Despite growing interest in KANs, their performance relative to MLPs on high-dimensional image classification tasks remains insufficiently characterized, partly because differences in model capacity and training protocols complicate direct comparisons. This study presents a parameter-controlled empirical comparison of MLPs and KANs on CIFAR-10 across five capacity tiers ranging from 64K to 4M parameters. The model configurations are matched to within 2% in parameter count, while the training protocol is held consistent to facilitate an assessment of architectural differences across capacity levels. The results show that KANs achieve higher mean classification accuracy than MLPs at low-to-medium capacities (T₁–T₄), with gains of 0.3–1.1 percentage points that decrease as model capacity increases. At the highest capacity (T₅), MLPs marginally outperform KANs by 0.3 percentage points. However, MLPs consistently deliver 3.0–5.3 times higher accuracy per GPU-hour, supported by a 3.1–5.4-fold training-throughput advantage and 18–27 fewer convergence epochs. Per-class results further reveal differences in the categories where each architecture performs relatively well, with MLPs showing marginal advantages in classes with lower intra-class variability and KANs exhibiting gains in visually complex, higher-variability classes. Overall, the findings suggest that the relative advantages of MLPs and KANs depend on model capacity and computational constraints: KANs offer modest accuracy gains at lower capacities, whereas MLPs provide greater computational efficiency across all evaluated tiers. These results support a complementary assessment of the two architectures under the specific experimental conditions considered.