Learning in the machine: To share or not to share?

Area of Science:

• Computational Neuroscience
• Computer Vision
• Artificial Intelligence

Background:

• Weight-sharing is a foundational principle in Convolutional Neural Networks (CNNs), contributing significantly to their success.
• This principle, however, is biologically implausible in physical neural systems like the brain, highlighting a key discrepancy.
• The necessity and degree of weight-sharing in artificial neural networks remain open questions.

Purpose of the Study:

• To investigate the necessity of weight-sharing in neural networks, particularly in computer vision applications.
• To explore alternative neural connection patterns inspired by biological neural circuitry.
• To evaluate the performance of networks that relax the weight-sharing assumption.

Main Methods:

• Conducted simulations by relaxing the weight-sharing assumption in standard neural network architectures.
• Explored the use of Free Convolutional Networks (FCNs) with variable neuron connection patterns.
• Utilized translationally augmented data, analogous to video data, for training.

Main Results:

• Weight-sharing, while pragmatic, is not strictly necessary for achieving high performance in computer vision tasks.
• Free Convolutional Networks achieved performance comparable to standard CNNs when trained with translationally augmented data.
• FCNs learned translationally invariant representations, effectively approximating weight-sharing under these conditions.

Conclusions:

• The findings suggest that alternative network architectures, like FCNs, can overcome the limitations of strict weight-sharing.
• Biological plausibility can inspire more flexible and potentially more efficient artificial neural network designs.
• Translationally invariant representations can be learned without explicit weight-sharing, offering a new perspective on network design.