Expressive architectures enhance interpretability of dynamics-based neural population models

Area of Science:

• Computational Neuroscience
• Machine Learning
• Dynamical Systems Theory

Background:

• Artificial neural networks (ANNs) are crucial for decoding neural activity and understanding biological computation.
• Accurate and low-dimensional latent dynamics are essential for interpretable ANN models of neural data.
• Existing recurrent neural network (RNN)-based ANNs struggle to infer true latent dynamics from neural variance.

Purpose of the Study:

• To evaluate sequential autoencoders (SAEs) with different dynamics models for recovering latent chaotic attractors from simulated neural data.
• To compare the performance of RNN-based SAEs versus neural ordinary differential equation (NODE)-based SAEs.
• To identify architectural features that enable accurate inference of low-dimensional neural dynamics.

Main Methods:

• Simulated neural datasets with known latent chaotic attractors were generated.
• Sequential autoencoders (SAEs) employing RNN-based dynamics and NODE-based dynamics were implemented and trained.
• Performance was assessed by the accuracy of inferred firing rates, latent state dimensionality, trajectory recovery, and fixed point structure.

Main Results:

• RNN-based SAEs failed to infer accurate firing rates at the true latent dimensionality and incorporated spurious dynamics.
• NODE-based SAEs successfully inferred accurate firing rates at the correct latent dimensionality, recovering key dynamical features.
• Ablation studies indicated that NODE's ability to use higher-capacity MLPs for vector fields and predict derivatives improved performance.

Conclusions:

• NODE-based SAEs offer superior interpretability and accuracy in recovering low-dimensional latent dynamics compared to RNN-based SAEs.
• The architecture of the dynamics model significantly impacts the ability to infer meaningful neural dynamics.
• NODE-based dynamics provide a promising alternative for modeling neural activity and advancing computational neuroscience.