Core Concepts
MotorNet is an open-source Python toolbox designed to address limitations in neural control of movement by providing a flexible, differentiable, and biomechanically realistic effector training platform using artificial neural networks.
Abstract
MotorNet is a Python toolbox developed to overcome challenges in neural control of movement by integrating differentiable, complex, and realistic effectors with artificial neural networks. The toolbox offers ease of installation, high-level user-friendly API, and modular architecture for model building. MotorNet allows training ANNs on motor tasks efficiently and effectively.
The content discusses the development and functionality of MotorNet, highlighting its key features such as ease of use, flexibility in model building, compatibility with PyTorch updates, and open-source nature. The tool aims to streamline the process of training ANNs on motor tasks by providing a comprehensive framework for researchers.
Key points include the importance of artificial neural networks (ANNs) in understanding brain function related to movement control. The limitations faced due to separate platforms for neural control and biomechanical simulation are addressed through the development of MotorNet. The toolbox allows efficient training on motor tasks using ANNs with realistic effectors that are differentiable.
The study also presents results from training an ANN on a centre-out reaching task against a curl field using the arm26 model within MotorNet. It demonstrates how the toolbox can be used to replicate established results from previous studies related to preferential movement direction tuning based on effector geometry.
Overall, MotorNet serves as a valuable tool for researchers working on neural control of movement by providing a comprehensive platform for training ANNs on motor tasks efficiently and realistically.
Stats
Training time: 13 minutes on Mac Studio with M1 Max CPU
Batch size: 64
Number of GRUs: 50 units
Quotes
"MotorNet is designed to meet several goals: ease of installation, ease of use, a high-level user-friendly API." - Content
"Training an ANN using backpropagation through itself enables fast and efficient learning." - Content