The Role of Architectural and Learning Constraints in Neural Network Models: A Case Study on Visual Space Coding

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Deep learning has revolutionized engineering but is underutilized in neurocomputational modeling.
• Unsupervised deep learning, focusing on generative learning, offers a promising avenue for advancing models of perception and cognition.

Purpose of the Study:

• To investigate the efficacy of unsupervised deep learning architectures for neurocomputational modeling.
• To compare generative (unsupervised) and discriminative (supervised) learning approaches in modeling neural coding of visual space.
• To assess how network architectures and learning rules influence emergent neural representations.

Main Methods:

• Simulations comparing Restricted Boltzmann Machines (RBMs) and autoencoders against supervised feed-forward networks.
• Systematic testing of receptive fields and gain modulation in hidden neurons.
• Exploration of sparse coding principles within unsupervised learning architectures.

Main Results:

• Architectural and learning constraints significantly shape the emergent coding of visual space.
• Unsupervised models, especially RBMs, demonstrated a closer alignment with neurophysiological data from primate parietal cortex.
• The study identified specific network properties relevant for modeling biological neural information processing.

Conclusions:

• Unsupervised deep learning, particularly RBMs, provides a more biologically plausible framework for neurocomputational models than traditional supervised methods.
• Generative learning principles are crucial for understanding neural coding in sensorimotor transformations.
• This research offers novel insights into the application of artificial neural network properties for modeling brain function.