Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
This project is dedicated to the implementation and research of Kolmogorov-Arnold convolutional networks. The repository includes implementations of 1D, 2D, and 3D convolutions with different kernels, ResNet-like and DenseNet-like models, training code based on accelerate/PyTorch, as well as scripts for experiments with CIFAR-10 and Tiny ImageNet.
| Date | Stars |
|---|---|
| 2026-07-30 | 531 |
| 2026-08-11 | 530 |
| 2026-08-18 | 529 |
| 2026-08-19 | 530 |
| 2026-09-07 | 529 |
| 2026-09-08 | 530 |
| 2026-09-14 | 529 |
| 2026-09-21 | 530 |
| 2026-10-03 | 531 |
| 2026-10-08 | 532 |
| 2026-10-17 | 531 |
| 2026-10-20 | 532 |
| 2026-10-25 | 531 |
| 2026-10-30 | 532 |
| 2026-11-04 | 532 |
Today
— stars today
This week
+1 stars this week
This month
+1 stars this month
Momentum
0.0
growth rate 0.19%/day
[](https://arxiv.org/abs/2407.01092) # TorchConv KAN: A Convolutional Kolmogorov-Arnold Networks Collection This project introduces and demonstrates the training, validation, and quantization of the Convolutional KAN model using PyTorch with CUDA acceleration. The `torch-conv-kan` evaluates performance on the MNIST, CIFAR, TinyImagenet and Imagenet1k datasets. ## Project Status: Under Development ### Updates - ✅ [2024/05/13] Convolutional KALN layers are available - ✅ [2024/05/14] Convolutional KAN and Fast KAN layers are available - ✅ [2024/05/15] Convolutional ChebyKAN is available now. MNIST, CIFAR10, and CIFAR100 benchmarks are added. - ✅ [2024/05/19] ResNet-like, U-net like and MoE-based (don't ask why=)) models released with accelerate-based training code. - ✅ [2024/05/21] VGG-like and DenseNet-like models released! Gram KAN convolutional layers added. - ✅ [2024/05/23] WavKAN convolutional layers added. Fixed a bug with the output hook in ```trainer.py```. - ✅ [2024/05/25] U2-net like models added. Fixed a memory leak in ```trainer.py```. - ✅ [2024/05/27] Updated implementation of WavKAN - much faster now. Added VGG-WavKAN. - ✅ [2024/05/31] Fixed KACN Conv instability issue, added Lion optimizer, updated baseline models and benchmarks, and :fire::fire::fire:**pretrained weights on Imagenet1k are released**:fire::fire::fire:, as well as Imagenet1k training scripts. - ✅ [2024/06/03] JacobiKAN Convs are available now. - ✅ [2024/06/05] BernsteinKANs and BernsteinKAN Convs are available now. - ✅ [2024/06/15] Introducing Bottleneck KAN Convs (with Gram polynomials as basis functions for now). Added LBFGS optimizer support (it's not well-tested, please raise an issue if you face any problems with it). Regularization benchmarks on CIFAR 100 are published. Hyperparameters tuning with **Ray Tune** are released. - ✅ [2024/06/18] ReLU KAN Convs are available now. - ✅ [2024/06/20] :fire::fire::fire:**New pretrained checkpoint on Imagenet1k are released**:fire::fire::fire: VGG11 style with Bottleneck Gram Convolutions. The model achieves 68.5% Top1 accuracy on Imagenet1k validation set with only 7.25M parameters. - ✅ [2024/07/02] :fire::fire::fire: We released our paper :fire::fire::fire: [Kolmogorov-Arnold Convolutions: Design Principles and Empirical Studies](https://arxiv.org/abs/2407.01092) - ✅ [2024/07/09] PEFT code for KAGN models are released, as well as new RDNet-like models, better implementation of Kolmogorov-Arnold-Gram concolutions, and medical images segmentation scripts. - ✅ [2026/08/17] Introducing ConvNeXtKAGNBlock — a ConvNeXt V2-style block with a Gram KAN layer in place of the MLP — and KAGNConv(1,2,3)DLayerV3 with built-in GRN. Plus ConvNeXtKAGN model. ### TODO list and next steps - Right now VGG19-like model is training on Imagenet1k - Right now Resnet50-like model is training on Imagenet1k - Finetuning experiments on other benchmarks are in progress, as well as PEFT methods exploration - I'm working on pruning and visualization methods as well --- ## Table of content: - [Introducing Convolutional KAN layers](#item-one) - [Model Zoo](#item-two) - [Performance Metrics](#item-three) - [Discussion](#item-four) - [Usage](#item-five) - [Accelerate-based training](#item-six) - [Contributions](#item-seven) - [Acknowledgements](#item-eight) - [References](#item-nine) <a id="item-one"></a> ## Introducing Convolutional KAN layers Kolmogorov-Arnold networks rely on Kolmogorov-Arnold representation theorem:  So, from this formula, the authors of [KAN: Kolmogorov-Arnold Networks](https://arxiv.org/abs/2404.19756) derived the new architecture: learnable activations on edges and summation on nodes. MLP in opposite performs fixed non-linearity on nodes and learnable linear projections on edges.  In
Excerpt of 19,804 characters
Read on GitHub
62
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:a2fb9705b1e3b281, topic:computer-vision