Scope

The main driving philosophy of nneuroutil is that it should be backend-agnostic. All current algorithms are implemented against the Array API Standard and work with both mutable backends (e.g. numpy or torch) and immutable backends (e.g. jax).

On top of that, the scope is not to overthrow other libraries. The algorithms that are added here should be either (a) standard and not worth an extra dependency or (b) novel / with different tradeoffs.

An example of this is the LDA algorithm implemented in nneuroutil.classifier, which uses a GSVD-based formulation for rank-deficient problems where the within-class scatter matrix is singular. Unlike scikit-learn’s default SVD-based LDA, it can retain discriminative directions in the null space of that matrix, which can be useful when there are more features than samples.

Install

pip install nneuroutil
# or: uv add nneuroutil

Optional backends are extras: nneuroutil[jax], nneuroutil[torch], and nneuroutil[visualization].

Quick Start

The library contains a few different algorithms and tools for general analysis of (mainly) timeseries data, classification, constructing neural network models, etc.

It’s generally pretty straightforward. For example, the code below shows a standard LDA classifier based on the GSVD algorithm of Howland & Park (2004) that uses jax.

import jax
import jax.numpy as jnp

from nneuroutil.classifier import classify_linear_discriminant_analysis

key = jax.random.PRNGKey(0)
key1, key2 = jax.random.split(key)
features = [
    jax.random.normal(key1, (50, 2)) * 0.5 + jnp.array([-2.0, 0.0]),
    jax.random.normal(key2, (50, 2)) * 0.5 + jnp.array([2.0, 0.0]),
]

clf = classify_linear_discriminant_analysis(features, [0, 1])
print(clf.predict(jnp.array([[1.8, 0.2], [-1.9, -0.1]])))  # [1. 0.]

Links

• Code.

• Documentation.

• License: MIT (see LICENSES/MIT.txt).