Constraints of Biological Neural Networks and Their Consideration in AI Applications

Table 1

A summary of constraints of biological neural networks and the possible associated constraints for designing artificial neural networks. For further examples and references of these constraints refer to the main text.

Process or structure	Constraints in biological networks	Possible constraints in artificial networks
Chemical synapses	Slow and prone to fatigue	Synaptic modification possible without these constraints or accurate representation of the biological process (although fatigue can be useful in learning)
Graded potential neurons	Electric charge (ions) leaks, so information can only be transported a short distance	Electronics or computer simulations do not present these constraints
Spiking neurons	Less information can be carried than with graded potential neurons	No need to simulate spiking neurons, as no artificial constraints on graded potential neurons exist
Size of neural networks	Energetically costly, especially in small/cognitively simple animals	Artificial networks will have higher energy costs. This may be a constraint depending on the application (i.e., importance of battery life)
Ecological function	Many well studied neural networks are related to survival behaviours—and are tightly tuned to these behaviours	Artificial networks designed to implement different behaviours may not work, as they are not performing exactly the same role
Multi-stimuli responses	No single neural network is optimal for a behaviour, and several environmental stimuli and neural pathways combine to produce a robust behaviour	Optimising a single neural network for a behaviour may be difficult. Unexpected (and possibly poor) results may occur in overly simple environments