A novel nonlinear adaptive filter using a pipelined second-order Volterra recurrent neural network

Area of Science:

• Computational Neuroscience
• Machine Learning
• Signal Processing

Background:

• Recurrent Neural Networks (RNNs) face challenges with heavy computational complexity and performance limitations.
• Existing nonlinear adaptive filters often struggle with efficiency and accuracy in complex signal processing tasks.

Purpose of the Study:

• To introduce a novel nonlinear adaptive filter, the pipelined second-order Volterra recurrent neural network (PSOVRNN), to enhance performance and reduce computational load.
• To detail a modified real-time recurrent learning (RTRL) algorithm for the proposed PSOVRNN.
• To evaluate the effectiveness of PSOVRNN in nonlinear colored signals prediction and nonlinear channel equalization.

Main Methods:

• The PSOVRNN is constructed using multiple small-scale second-order Volterra recurrent neural network (SOVRNN) modules.
• Modules are processed in parallel using pipelined parallelism for improved computational efficiency.
• Nonlinearity is introduced via recursive second-order Volterra (RSOV) expansion within each SOVRNN module.

Main Results:

• Computer simulations show PSOVRNN outperforms standard RNN and pipelined recurrent neural networks (PRNN) in nonlinear colored signals prediction.
• PSOVRNN demonstrates superior performance in nonlinear channel equalization compared to PRNN and RNN.
• The enhanced performance of PSOVRNN comes with increased computational complexity due to nonlinear expansion.

Conclusions:

• The proposed PSOVRNN offers a significant improvement in performance for nonlinear signal processing tasks.
• Pipelined parallelism and second-order Volterra expansion contribute to the enhanced capabilities of PSOVRNN.
• PSOVRNN represents a promising advancement in adaptive filtering, balancing performance gains with computational considerations.