ngc-learn: Building predictive coding and spiking networks on JAX

If you want to test whether a learning rule inspired by the brain can do work that backpropagation does differently — not just faster, but with different constraints on locality, timing, and energy — you need more than a standard deep learning framework. That is the niche ngc-learn fills. Maintained by the Neural Adaptive Computing (NAC) lab, it is a Python library for building, simulating, and analyzing biophysical and neurobiological systems, spiking neuronal networks, predictive coding circuitry, and biomimetic NeuroAI agents that learn in a biologically plausible manner. The project is built on JAX and distributed under the 3-Clause BSD license, which matters for teams that need to reuse code in commercial research or product prototypes without copyleft obligations. The maintainers point to official documentation and tutorials, a companion ngc-museum repository that implements several historical models, and a related paper that appeared in the Nature Neuromorphic Hardware and Computing Collection in 2023 and was selected as an Editors’ Highlight for Applied Physics and Mathematics in 2022. That provenance gives the library academic grounding, but for builders the more immediate question is what it costs to run. The stack you have to adopt Current releases, starting with 3.2.3, require Python 3.12 or newer, NumPy, SciPy, JAX, and ngcsimlib 3.1.1 or later. Plotting and density estimation add Matplotlib, imageio, and Scikit-learn. Installation is a single pip command — pip install ngclearn — but the default package pulls the CPU version of JAX if no GPU build is present. If you want GPU execution, you need to install the CUDA 12 variant of JAX and Jaxlib first. That ordering is not a footnote. The README warns that with CUDA 12 and JAX versions above 0.4.28, convolution and deconvolution routines can fail to compile on older Pascal-generation GPUs (compute capability 6.1, such as