A unit paradox for artificial neuronal networks

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Formal and artificial neuronal networks (ANN) traditionally use unit step functions.
• Recent trends involve replacing step functions with S-shaped curves, often with adaptive learning procedures.
• The dynamical consequences of this substitution on single unit discharges remain underexplored.

Purpose of the Study:

• To investigate the dynamical consequences of replacing step functions with S-shaped curves in artificial neuronal networks.
• To address the 'unit paradox' where model units partially correspond to real neural units.
• To propose resolutions for the apparent contradiction between artificial and biological neural network models.

Main Methods:

• Analysis of artificial neuronal network (ANN) models with S-shaped activation functions.
• Examination of the time course of single unit discharges in these models.
• Theoretical propositions to reconcile discrepancies between artificial and biological neural units.

Main Results:

• The dynamical implications of using S-shaped curves in ANNs require further examination.
• A 'unit paradox' exists due to partial correspondence between model and real neural units.
• Propositions are presented to resolve the unit paradox and related contradictions.

Conclusions:

• Revisiting the relationship between different trends in neural network theory is necessary.
• The substitution of step functions with S-shaped curves in ANNs has underexamined dynamical consequences.
• Resolving the unit paradox is crucial for a more accurate understanding of ANN models.