pAMICA

Python (PyTorch) implementation of Adaptive Mixture Independent Component Analysis (AMICA) that reproduces the results of the reference Fortran binary, with GPU / Apple-GPU / CPU support. AMICA is a blind source separation algorithm widely used for electroencephalography (EEG) and electromyography (EMG) source decomposition.

pamica exposes a scikit-learn-style interface over a natural-gradient expectation-maximization (EM) backend that matches the Fortran reference (amica15) to within numerical tolerance: single-model log-likelihood and Hungarian-matched component correlation both agree with Fortran on real EEG.

Why pamica

• Fortran parity is the specification. Correctness is defined as matching the reference Fortran output within tolerance, not merely converging. See Validation & Parity.
• Multiple backends, one API. The default PyTorch natural-gradient EM backend runs on CUDA, CPU, and (in float32) Apple MPS; an optional MLX backend targets Apple-Silicon GPUs, and a legacy NumPy reference is retained as an oracle. See Backends & Devices.
• Feature-complete against the reference. Newton optimization, exact-EM mixture updates, all five source-density families, component sharing, and outlier rejection are ported and validated.

Quick links

• Getting Started: install and run your first decomposition.
• Backends & Devices: pick CUDA / CPU / MLX and float32 vs float64.
• Validation & Parity: how correctness is defined and checked.
• pamica vs. AMICA: every deliberate difference from the reference.
• API Reference: the AMICA interface and the backend classes.

Precision and parity

The natural-gradient backend computes in float64 for Fortran parity. Apple MPS cannot represent float64, so parity runs use CPU or CUDA. float32 is required on MPS/MLX and carries about seven significant digits; it runs about as fast as float64 on CUDA and modestly faster on CPU, and the Apple-GPU speed comes from the MLX backend. Use float64 for reference-parity runs.