A fast C++ / CUDA engine for training neural networks with predictive coding, a brain-inspired alternative to backpropagation with a simple Python interface.

Docs • Source • PyPI • Contributing

Most neural networks learn with backpropagation: one global error signal flows backward through every layer. Predictive coding networks (PCNs) learn differently — each layer only ever reacts to the mismatch between its own activity and a prediction from the layer next to it, a theory with roots in neuroscience for how the brain itself learns.

Deepity is a from-scratch engine for training PCNs quickly: a hand-tuned C++20 core (AVX2/AVX-512 SIMD, OpenBLAS/MKL) with an active CUDA backend, and a Python package (pydeepity) on top so you can define and train a network in a few lines.

pip install pydeepity

from pydeepity import SimplePCN, Linear, TanH

net = SimplePCN(Linear(2, 8), TanH(), Linear(8, 1), batch_size=1)
net.configure(learning_rate=0.05, inference_rate=0.3, optimizer="SGD")

On MNIST, Deepity's SimplePCN reaches 98.21% test accuracy in 487.5s on a laptop CPU — faster than the two established predictive-coding libraries it was benchmarked against (JAX-based jpc, and ngc-learn) at matching or better accuracy. See the full benchmark writeup for the exact setup and where Deepity does not win.

Deepity started as an academic research project and is now becoming infrastructure other predictive-coding research can build on. The CPU path is the mature one; the CUDA backend is under active development, with large-scale ImageNet training as the current focus.

We're looking for contributors with C++/CUDA performance or Python build/DevOps experience: see the main repo's open issues.

_{MIT License}