Reproduce “Kolmogorov-Arnold Network Autoencoders” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mohammadamin Moradi, Shirin Panahi, Erik Bollt, Ying-Cheng Lai
Kolmogorov-Arnold Network (KAN) autoencoders demonstrate competitive to superior reconstruction accuracy and latent feature quality compared to Convolutional Neural Network (CNN) autoencoders on image datasets, notably outperforming CNNs on SVHN and CIFAR-10. These KAN-based models, while having a higher parameter count, maintain comparable training times due to an efficient implementation.
Key insights
- KANs' unique architecture, placing flexible spline-based activation functions directly on network edges, enables them to effectively learn efficient and discriminative latent representations.
- The inherent design of KANs provides greater representational capacity, allowing them to capture a broad range of complex, non-linear dependencies within data, covering both local and global patterns.
- Despite KAN autoencoders often having a substantially higher number of trainable parameters, optimized implementations can mitigate computational bottlenecks, leading to training times comparable to CNN autoencoders.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.