Research map / Learning theory and algorithms / Prediction and decision losses

Normative Networks for Source Separation via Local Plasticity and Dendritic Computation

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Theory of neural network sensitivity · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Predictive Entropy Maximization (PEM), an online, local-plasticity neural network for structured blind source separation that replaces the exact log-determinant (correlative entropy) objective with a second-order Taylor surrogate.

Problem

Online, biologically plausible blind source separation of structured (domain-constrained) and correlated sources; existing determinant-based online networks (CorInfoMax) need inverse-covariance states and nonlocal recurrent dynamics.

Approach

Taylor-expand the regularized log-determinant of the output covariance around its diagonal, yielding a surrogate with a variance-expansion term and a variance-normalized cross-covariance penalty; this gives gradients in covariance traces (no inverse states), so feedforward weights follow an error-driven rule, lateral inhibition is Hebbian, and domain constraints are output nonlinearities/projectio

Key results

• PEM and u-PEM remain robust as source correlation rho increases to 0.5 and outperform ICA-InfoMax and NSM, whose performance degrades rapidly.
• For sparse sources across input SNR 5-30 dB, PEM and u-PEM track the batch CorInfoMax (Batch) oracle closely.
• Auditory separation on 30 random mixing scenarios yields source-wise SNRs of 24.12±1.98, 25.07±2.14, and 21.54±2.17 dB (95% CI).

Surprising findings

• PEM is more resistant to source correlation than its unnormalized variant u-PEM, suggesting variance normalization helps under correlated sources.

Limitations

• Performance is sensitive to hyperparameter choices and to the number of observed mixtures.
• Computational cost scales with the number of fast inference iterations.
• Surrogate approximation error grows with source correlation (though bounded).

Experimental setup

Datasets: synthetic copula-t and uniform source mixtures pre-whitened 12x12 natural image patches (Sparsenet) librosa audio clips (fishin, pistachio, vibeace)

Benchmarks: synthetic BSS mixtures (mean component SNR) natural image patch sparse receptive fields auditory source separation

Metrics: mean component SNR (mSNR) source-wise SNR Taylor remainder vs bound

Baselines: CorInfoMax (Batch) ICA-InfoMax CorInfoMax (Online) Nonnegative Similarity Matching (Online)

Builds on: CorInfoMax ICA-InfoMax Nonnegative Similarity Matching predictive coding VICReg

Closest papers by meaning

• Correlative Information Maximization Based Biologically Plausible Neural Networks for Correlated Source Separation (ICLR 2023 · similarity 0.89)
• Complex priors and flexible inference in recurrent circuits with dendritic nonlinearities (ICLR 2024 · similarity 0.86)
• Divide-and-Conquer Predictive Coding: a structured Bayesian inference algorithm (NeurIPS 2024 · similarity 0.85)
• Neurons as Detectors of Coherent Sets in Sensory Dynamics (NeurIPS 2025 · similarity 0.85)
• Dendritic Localized Learning: Toward Biologically Plausible Algorithm (ICML 2025 · similarity 0.85)
• Decomposing stimulus-specific sensory neural information via diffusion models (NeurIPS 2025 · similarity 0.85)