Interleaving cortex-analog mixing improves deep non-negative matrix factorization networks

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Deep learning

Background:

• Biological constraints improve artificial neural network (ANN) performance.
• The positive nature of long-range cortical signals has not previously improved ANN performance.
• Non-negative matrix factorization (NMF) models positive long-range interactions but deep convolutional neural networks (CNNs) with NMF modules underperform.

Purpose of the Study:

• To investigate if incorporating positive long-range signaling, analogous to cortical processing, enhances deep neural network performance.
• To determine if combining NMF's positive activities in intermediate modules improves upon conventional CNNs.

Main Methods:

• Developed novel deep convolutional neural network modules integrating Non-negative matrix factorization (NMF) principles.
• Introduced intermediate modules that combine positive activities, mimicking cortical column processing.
• Evaluated network performance on benchmark datasets against standard deep convolutional networks.

Main Results:

• The proposed network architecture, incorporating positive long-range signaling and local interactions, significantly improved performance on benchmark data.
• Performance exceeded that of conventional deep convolutional neural networks (CNNs) of comparable size.
• The findings demonstrate the benefit of mimicking cortical hyper-column processing in ANNs.

Conclusions:

• Integrating positive long-range signaling with local interactions, inspired by cortical hyper-columns, enhances deep network performance.
• This biologically inspired approach offers a promising avenue for developing more powerful and efficient deep learning models.
• The study validates the potential of incorporating specific biological signaling mechanisms into artificial neural networks.