NeuroTabModels.jl
Deep learning for tabular data, compiled to XLA with Reactant.jl and differentiated with Enzyme.
- Organisation
- Evovest
- Period
- 2026 – present
- Stack
- Julia, Lux.jl, Reactant.jl, Enzyme, Zygote
- Code
- Evovest/NeuroTabModels.jl
NeuroTabModels is Evovest’s library of neural networks for tabular data. I rebuilt its training stack and added recent architectures from the tabular deep learning literature.
What I did
- Compiled training with Reactant.jl. I moved the models to Lux.jl and compiled training and inference to XLA through Reactant.jl, instead of running them op by op from Julia.
- Multiple AD backends. Gradients come from Enzyme or Zygote, on CPU, GPU or Reactant, each loaded only when the user asks for it.
- New architectures. TabM in all three of its variants, ModernNCA, numerical embeddings shared by every model, and temporal embeddings for time-dependent data.
- ModernNCA at scale. ModernNCA attends over the whole training set, so memory grew with the data. I rewrote its attention to work in chunks with a FlashAttention-style backward pass, so memory stays flat no matter how large the training set is (in review).