PsychRNN

PsychRNN trains recurrent neural networks to model cognitive tasks and investigate neural circuit mechanisms in computational, systems, and cognitive neuroscience.

Key Features:

• TensorFlow training backend: Uses TensorFlow as the training backend for optimizing recurrent neural network models.
• Python and NumPy model definition: Enables task and model specification using Python and NumPy-based interfaces.
• Extensibility: Allows customization and extension of models and training procedures for users with TensorFlow expertise.
• Neurobiologically relevant constraints: Supports imposition of constraints on synaptic connectivity patterns to reflect neurobiological structure.
• Modular task specification: Provides a modular task framework that permits parametric variation in task demands.
• Curriculum learning (task shaping): Implements curriculum learning with performance-based adjustments to task difficulty during training.
• Post-training analysis support: Facilitates analysis of network properties and dynamics such as synaptic connectivity patterns and neural activity.

Scientific Applications:

• Modeling cognitive computations: Train RNNs to perform cognitive tasks to investigate neural representations and circuit mechanisms underlying cognition.
• Cross-species task comparison: Fit networks on tasks used in animal and human experiments to compare artificial and biological task behaviors.
• Connectivity studies: Examine how imposed synaptic connectivity constraints influence network solutions and computations.
• Neural population dynamics: Analyze dynamics of neural activity during task execution and how population activity supports behavior.
• Perturbation and learning trajectories: Conduct perturbation studies and investigate how curriculum/task shaping trajectories affect learning outcomes.

Methodology:

Define a cognitive task in Python/NumPy, train a recurrent neural network using TensorFlow, optionally impose neurobiologically informed synaptic connectivity constraints during training, apply curriculum learning for iterative training adjustments, and perform post-training analyses of synaptic connectivity and neural activity dynamics.