factorized-lowrank-rnn-independent-latent

name	factorized-lowrank-rnn-independent-latent
description	Factorized Low-Rank RNN (FacRNN) framework for uncovering independent neural latent dynamics and connectivity. Group-wise independence among latent dimensions with variational autoencoder formulation and partial correlation penalty. Disentangles interpretable latent trajectories in low-dimensional space for neural population activity analysis. Use for: neural latent dynamics discovery, low-rank connectivity interpretation, disentangled representation learning, neural population modeling, independent dimension analysis, VAE-based RNN. Activation: factorized RNN, low-rank RNN, independent latent, disentangled dynamics, group-wise independence, partial correlation, neural population, latent trajectory, interpretable connectivity.
license	Complete terms in LICENSE.txt
metadata	{"arxiv_id":"2511.13899","published":"2026-06-02","authors":"Chengrui Li, Yunmiao Wang, Yule Wang, Weihan Li, Dieter Jaeger, Anqi Wu","tags":["low-rank-rnn","latent-dynamics","disentanglement","variational-autoencoder","neural-population","connectivity","interpretability"]}

Factorized Low-Rank RNN for Independent Neural Latent Dynamics

FacRNN framework that uncovers independent neural latent dynamics through group-wise independence assumptions and variational autoencoder formulation.

Core Problem

Standard low-rank RNNs (lrRNNs) uncover low-dimensional latent dynamics from neural population activity, but their functional connectivity lacks independence interpretations — making it difficult to assign distinct computational roles to different latent dimensions.

Solution: Factored Recurrent Neural Network (FacRNN)

Key Innovation

Group-wise independence among latent dimensions while allowing flexible within-group entanglement.

• Independent latent groups: Dynamics evolve separately
• Within-group richness: Complex computation possible within groups
• Interpretability: Clear computational roles for each group

Methodology

Variational Autoencoder Framework

Reformulate lrRNN under VAE framework to introduce partial correlation penalty that encourages independence between groups of latent dimensions.

Latent Groups: {G1, G2, ..., Gk}
Each group: Independent evolution
Within-group: Flexible entanglement

Partial Correlation Penalty

• Encourages independence: Between latent groups
• Preserves flexibility: Within-group dynamics remain rich
• Improves disentanglement: Clear dimension separation

Architecture Components

1. Encoder: Maps neural observations to latent groups
2. Recurrent dynamics: Independent evolution per group
3. Decoder: Reconstructs neural activity from groups
4. Independence penalty: Partial correlation regularization

Experimental Validation

Synthetic Data

• Improved disentanglement: Clearer latent dimension separation
• Baseline comparison: Superior to standard lrRNN

Monkey M1 Data

• Motor cortex: Interpretable latent dynamics for movement
• Connectivity clarity: Independent groups match functional roles

Mouse Voltage Imaging

• Large-scale neural activity: Effective disentanglement
• Trajectory interpretability: Clear latent structure

Key Advantages

Feature	Standard lrRNN	FacRNN
Independence	Implicit	Explicit group-wise
Interpretability	Limited	Clear dimension roles
Disentanglement	Weak	Strong with penalty
Within-group complexity	Same	Preserved
Latent trajectory clarity	Ambiguous	Separable groups

Applications

Neural Population Analysis

• Latent dynamics discovery: Clear dimension separation
• Connectivity interpretation: Independent group roles
• Computational role assignment: Distinct functions per group

Motor Cortex Modeling

• M1 dynamics: Movement-related latent groups
• Independent movement components: Separate trajectory control
• Connectivity interpretation: Motor function roles

Voltage Imaging Analysis

• Large-scale data: Effective group separation
• Trajectory interpretation: Clear latent evolution
• Connectivity patterns: Independent group structure

Implementation Patterns

Group Organization

# Latent dimension grouping
latent_dims = 10
groups = [
    [0, 1, 2],   # Group 1: Motor planning
    [3, 4, 5],   # Group 2: Execution
    [6, 7, 8, 9] # Group 3: Feedback
]

# Each group evolves independently
for group in groups:
    group_dynamics = rnn_group_step(group, input)

Partial Correlation Penalty

• Cross-group: Minimize correlation between groups
• Within-group: Allow flexible interactions
• Regularization: Partial correlation in loss function

VAE Training

Loss = Reconstruction + KL Divergence + Independence Penalty

Independence penalty: Partial correlation between group outputs

Relation to Neuroscience

Neural Population Dynamics

• Latent factors: Match neural computation components
• Independent processes: Separate cognitive functions
• Within-group integration: Complex local processing

Connectivity Interpretation

• Functional roles: Clear assignment to latent groups
• Independent circuits: Separate network modules
• Integrated computation: Within-group cooperation

Motor Control

• M1 organization: Movement decomposition matches groups
• Planning/execution: Independent latent processes
• Feedback integration: Separate group for adaptation

Pitfalls

Group Assignment

• Manual grouping: Requires domain knowledge
• Group number: Optimal split unclear
• Dimension allocation: Wrong assignment disrupts independence

Training Challenges

• Penalty strength: Too strong kills within-group dynamics
• KL balance: Trade-off with reconstruction quality
• Group collapse: Underutilized groups may die

Interpretation Issues

• Role assignment: Needs neuroscience validation
• Group meaning: Semantic interpretation required
• Cross-group interaction: Unexpected dependencies may persist

Comparison with Related Methods

Standard Low-Rank RNN

• No independence: All dimensions entangled
• Limited interpretability: Ambiguous roles
• FacRNN advantage: Explicit group separation

Independent Component Analysis

• Complete independence: No within-group flexibility
• FacRNN difference: Group-wise with internal richness
• Trade-off: Balance isolation and complexity

Dynamical Systems Decomposition

• Decoupling methods: Similar independence goals
• FacRNN innovation: VAE framework + partial correlation
• Advantage: Learned disentanglement with reconstruction

Activation Keywords

• factorized low-rank RNN
• FacRNN
• independent latent dynamics
• group-wise independence
• partial correlation penalty
• disentangled representation
• neural population latent
• interpretable connectivity
• VAE RNN
• latent trajectory separation

Related Skills

• neural-dynamics-universal-translator-foundation: Universal neural dynamics
• behavior-decomposed-lds: Behavior decomposition in LDS
• neural-manifold-learning-dynamics: Manifold learning for dynamics
• low-rank-rnn-learning-dynamics: Learning dynamics in low-rank RNNs
• neural-population-decoding: Population decoding methods

References

• arXiv:2511.13899 - Factorized Low-Rank RNN Framework (Li et al., 2026)
• Low-rank RNN theory literature
• Variational autoencoder frameworks
• Neural population dynamics studies
• Disentangled representation learning