Aligned and oblique dynamics in recurrent neural networks.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience

Background:

• The relationship between neural activity and behavior is fundamental to neuroscience.
• Partial dissociations between neural activity and external variables suggest complex internal dynamics.
• A theoretical framework explaining these relationships is currently lacking.

Purpose of the Study:

• To explore the geometrical relationship between neural dynamics and network output using recurrent neural networks (RNNs).
• To understand the conditions and mechanisms governing neural representations and their potential dissociations.

Main Methods:

• Utilized recurrent neural networks (RNNs) to model neural dynamics.
• Analyzed network outputs from geometrical and dynamical stability perspectives.
• Investigated the role of readout weight magnitude as a control parameter.
• Examined neural recordings for evidence of distinct dynamical regimes.

Main Results:

• Identified two dynamical regimes in RNNs: aligned and oblique.
• Demonstrated that readout weight magnitude controls the transition between regimes.
• Showed that oblique networks exhibit greater heterogeneity, noise suppression, and robustness.
• Found the oblique regime to be specific to recurrent networks due to dynamical stability.
• Observed dissociations in neural recordings corresponding to these regimes.

Conclusions:

• The study provides a new geometrical framework for understanding neural representations.
• Findings suggest that the oblique regime in RNNs offers advantages for information processing and robustness.
• The results offer a novel perspective for interpreting neural activity in relation to network dynamics and behavior.