Gradient-free training of recurrent neural networks using random perturbations

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Recurrent Neural Networks (RNNs) are powerful for sequential data but face training challenges.
• Backpropagation Through Time (BPTT) is standard but has limitations like vanishing gradients and memory issues.
• Perturbation-based methods offer a simpler alternative but often suffer from inefficient optimization.

Purpose of the Study:

• To develop a novel perturbation-based learning method for RNNs that overcomes the limitations of existing approaches.
• To enhance the efficiency and generalization capabilities of RNN training using a new perturbation technique.
• To demonstrate that this new method is competitive with gradient-based methods like BPTT.

Main Methods:

• Extended the Activity-Based Node Perturbation (ANP) method to operate in the time domain for RNNs.
• Employed a perturbation-based learning strategy requiring only forward passes and a global reinforcement signal.
• Conducted experiments to validate the performance, convergence, and scalability of the proposed method.

Main Results:

• The proposed ANP-based method achieved performance competitive with Backpropagation Through Time (BPTT).
• The new method demonstrated similar convergence times and scalability compared to BPTT.
• Outperformed standard node perturbation and weight perturbation methods in RNN training.
• Showcased efficient learning and improved generalization capabilities.

Conclusions:

• Perturbation-based learning offers a viable and versatile alternative to gradient-based methods for training RNNs.
• The proposed time-domain ANP method provides an efficient and effective approach to RNN training.
• This method is well-suited for applications in neuromorphic computing due to its inherent advantages.