Backpropagation With N -D Vector-Valued Neurons Using Arbitrary Bilinear Products.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Traditional neural networks (NNs) process data as vectors of scalars, potentially missing associations among adjacent data points.
• Vector-valued neural learning is an emerging area in deep learning research.

Purpose of the Study:

• To introduce a novel vector neural architecture, the Arbitrary BIlinear Product NN (ABIPNN), designed to model associations within vector data.
• To demonstrate the efficacy of ABIPNN in complex signal processing tasks.

Main Methods:

• Developed the Arbitrary BIlinear Product NN (ABIPNN) architecture, where neurons process information as vectors.
• Utilized arbitrary bilinear products, including circular convolution and 7-D vector products, for feedforward projections.
• Applied ABIPNN to multispectral image denoising and singing voice separation.

Main Results:

• ABIPNN demonstrated substantial performance improvements over conventional NNs in applied tasks.
• The network successfully learned associations within the vector-valued data during training.
• Significant enhancements were observed in multispectral image denoising and singing voice separation.

Conclusions:

• The proposed ABIPNN architecture offers a powerful new approach for vector-valued neural learning.
• Modeling associations within vector data using bilinear products leads to improved performance in signal processing.
• ABIPNN shows promise for various applications requiring nuanced data representation.