tension: A Python package for FORCE learning.

Area of Science:

• Computational neuroscience
• Machine learning
• Neural network modeling

Background:

• First-Order, Reduced and Controlled Error (FORCE) learning is effective for training chaotic recurrent neural networks (RNNs), often surpassing gradient-based methods.
• A lack of standardized software frameworks hinders the widespread adoption and application of FORCE learning.
• Existing FORCE implementations may lack efficiency or ease of use for diverse network types.

Purpose of the Study:

• To introduce "tension," an open-source Python package providing a unified TensorFlow/Keras API for FORCE learning.
• To demonstrate the package's versatility in training various RNN architectures, including rate, spiking, and biologically constrained networks.
• To offer a simplified and efficient tool for researchers to implement and iterate on FORCE training for chaotic RNNs.

Main Methods:

• Development of an object-oriented Python package, "tension," leveraging TensorFlow/Keras.
• Implementation of a high-level, extensible API to accommodate different network types.
• Benchmarking "tension" against conventional RNNs and existing FORCE implementations.

Main Results:

• The "tension" package successfully trains rate networks, spiking networks, and biologically constrained networks through a shared API.
• FORCE training using "tension" achieved lower loss compared to conventional RNNs.
• The "tension" implementation demonstrated superior runtime performance over published FORCE implementations.

Conclusions:

• "Tension" provides an accessible and efficient software framework for FORCE learning in chaotic RNNs.
• The package simplifies the process of training and iterating on complex neural network models.
• This work facilitates the study of emergent behaviors in neural dynamics through advanced RNN training techniques.