Vulpes: A functional approach to deep machine learning on the GPU

on YouTube ↗44m

" summary

Open source Vulpes is written in F# and uses QuantAlea's CUDA monad to push matrix work to the GPU, composing kernels so launches stay cheap and memory copies are minimized. It runs a deep belief net on MNIST and reports around 4% error on that dataset, with greedy layerwise pretraining plus fine‑tuned backprop and momentum. Code shows concrete kernel/launch wiring, gather/scatter for reporting, and works end‑to‑end on GPU.

AI-written summary · about

" chapters (19)

1. 0:00Intro
2. 3:46MNIST
3. 7:27potted history
4. 9:00Convergence
5. 9:49Backpropagation
6. 14:15Hidden Units
7. 17:56Belief Nuts
8. 18:39Layered Learning
9. 21:35Restricted Boltzmann Machine
10. 23:36Weight Updates
11. 24:48Momentum
12. 26:31Vulpas
13. 27:29Matrix Multiplication
14. 28:29CUDA C
15. 29:34CUDA F
16. 31:28CUDA Monad
17. 32:42CUDA Monad Example
18. 35:11Code Example
19. 38:22Questions

" related talks

• 30mAnyone can Build Great Deep Learning Applications - Deep NumpyCyrus Vahid · PyData Warsaw 2019
• 26mEfficiently serving LLMs at scale.Nick Hill · Scale By the Bay 2023
• 57mDeep Learning with C++Peter Goldsborough · Meeting C++ 2017
• 24mLevanter: Legible, Scalable, Reproducible Foundation Models with JAXDavid Hall · Scale By the Bay 2023
• 31mThe Evolution RevolutionRobert Plummer · ML Conf EU 2020
• 13mHow Open-Source vLLM Topped the Artificial Analysis LeaderboardDebarshi Raha · Ray Summit 2026
• 29mBeyond CUDA: GPU Accelerated Computing on Cross-Vendor Graphics Cards with Vulk...Alejandro Saucedo · KubeCon + CloudNativeCon + Open Source Summit China 2021
• 38mHow to run Neural Nets on GPUs'Melanie Warrick · Strange Loop 2015