neural-encoder neural-encoder

neural-encoder

NeurIPS 2026 Python 3.12+ PyPI version Run Tests Documentation Status License GitHub

neural-encoder learns embeddings from repeated measurements and multiview data. It was designed for single-trial fMRI responses and other high-dimensional, noisy measurements in which several observations correspond to the same underlying sample.

The encoder learns the signal shared across repetitions or views and maps each measurement to a low-dimensional representation. Repeated measurements are used during training; after fitting, individual measurements can be encoded independently.

The method is based on Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry. It combines feature reliability weighting, PCA, cross-view ridge denoising, distilled multiset canonical correlation analysis (MCCA), reliability weighting of the embedding dimensions, and nonlinear residual refinement.

Quick start

Fit from a matrix containing all repeated measurements. sample_ids gives the identity of the underlying sample represented by each row:

from neural_encoder import NeuralEncoder

encoder = NeuralEncoder()
Z_train = encoder.fit_transform(X_train, sample_ids=sample_ids)
Z_test = encoder.transform(X_test)

You can also fit from separate, aligned views. Corresponding rows of X1, X2, and X3 must describe the same sample:

encoder = NeuralEncoder()
encoder.fit_views([X1, X2, X3])

Z1, Z2, Z3 = (encoder.transform(X) for X in (X1, X2, X3))

See Usage for preprocessing, configuration, evaluation, and PyTorch export.

Indices