Feature ELMExtreme Learning Machines on CPU and CUDA

Composable feature maps, closed-form ridge and recursive least-squares solvers, and online and hierarchical ELM variants in modern C++20, with cuBLAS/cuSOLVER acceleration.

Get started

Choose a model

Try the demo

GitHub

⚡

Train in one solve

Random hidden layers and a single regularised least-squares solve. No back-propagation, no epochs.

🌊

Online and drifting streams

OS-ELM, ReOS-ELM, FOS-ELM and OS-CELM update chunk by chunk with recursive least squares and forgetting factors.

🧱

Hierarchical features

ELM auto-encoder layers stack into ML-ELM and H-OS-ELM for learned multilayer representations.

🟩

CUDA backend

cuBLAS GEMM hidden-layer transforms, a cuSOLVER Cholesky ridge solve and device-resident online RLS behind a single Backend::kGpu switch.

📊

Measured, not claimed

Google Benchmark suites and a digits-classification demo report accuracy and CPU/GPU timings side by side.

🐳

Docker-first

One dev image with CUDA 13.4, GoogleTest and Google Benchmark. CPU and GPU demo images that are ready for Hugging Face Spaces.

Tech stack

License and citation

License: MIT License
Citation guide: how to cite this project and the underlying algorithms

References

Huang, Guang-Bin, Qin-Yu Zhu, and Chee-Kheong Siew. 2006. Extreme learning machine: theory and applications.
Liang, Nan-Ying, Guang-Bin Huang, P. Saratchandran, and N. Sundararajan. 2006. A fast and accurate online sequential learning algorithm for feedforward networks.
Kasun, L. L. C., Yang Yang, Guang-Bin Huang, and Zhiping Zhou. 2013. Extreme learning machine for multilayer perceptron and autoencoder feature learning.
Broomhead, David S., and David Lowe. 1988. Multivariable functional interpolation and adaptive networks.
Moody, John, and Christian J. Darken. 1989. Fast learning in networks of locally-tuned processing units.