AngoraPy: A Python toolkit for modeling anthropomorphic goal-driven sensorimotor systems.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Robotics

Background:

• Deep learning models offer advantages over classical approaches in neuroscience by autonomously learning neural connectivity.
• Goal-driven models can generate testable hypotheses about brain function grounded in anatomical data.
• Applying goal-driven deep learning to sensorimotor systems is challenging due to the complexity of sensation-action loop training.

Purpose of the Study:

• To introduce AngoraPy, a Python library designed to simplify the training of recurrent convolutional neural networks for sensorimotor modeling.
• To provide researchers with accessible tools for goal-driven deep learning in computational neuroscience.
• To overcome the methodological hurdles in modeling the closed sensation-action loop.

Main Methods:

• Development of the AngoraPy Python library for training deep neural networks.
• Utilizing recurrent convolutional neural networks to model the human sensorimotor system.
• Training a recurrent toy model on an in-hand object manipulation task for illustration.

Main Results:

• AngoraPy successfully mitigates the complexity of training sensorimotor deep learning models.
• An illustrative example demonstrates the library's utility for in-hand object manipulation.
• Extensive benchmarks confirm AngoraPy's applicability across diverse control tasks (classical, 3D robotic, anthropomorphic).

Conclusions:

• AngoraPy empowers researchers to conduct goal-driven sensorimotor modeling with greater ease.
• The library's flexibility and adaptability support custom neural network architectures.
• AngoraPy significantly advances the application of deep learning in computational sensorimotor neuroscience.