Flow-field inference from neural data using deep recurrent networks

Area of Science:

• Computational neuroscience
• Machine learning applied to neurobiology

Background:

• Understanding neural population dynamics is crucial for deciphering brain computations like decision-making.
• Estimating these complex dynamics from neural data presents a significant methodological challenge.

Approach:

• Introduced Flow-field Inference from Neural Data using deep Recurrent networks (FINDR), an unsupervised deep learning technique.
• FINDR infers low-dimensional nonlinear stochastic dynamics from neural population activity.
• Applied FINDR to population spike train data from rat frontal cortex during an auditory decision-making task.

Key Points:

• FINDR surpasses existing methods in capturing heterogeneous neural responses.
• The method discovers interpretable low-dimensional dynamics by disentangling task-relevant and irrelevant neural activity.
• Learned dynamics facilitate visualization of flow fields and attractor structures.

Conclusions:

• FINDR offers a powerful approach for uncovering low-dimensional, task-relevant neural dynamics.
• This method advances the study of neural population activity and its role in cognitive computations.