Neural networks simulating short-term memory of two inputs with varying commonality

Area of Science:

• Computational neuroscience
• Cognitive neuroscience
• Neural network modeling

Background:

• Primate brain neural activity and connectivity underlying behavior are complex.
• Short-term memory in monkeys may involve sustained neural activity.
• Dynamic neural networks with sample-and-hold (SAH) functions model single-variable short-term memory tasks.

Purpose of the Study:

• Extend SAH network models to compute memory for two continuous-variable inputs.
• Investigate computational mechanisms of associative short-term memory for sensory signals with common mode components.
• Examine attractor states in SAH networks and their determinants.

Main Methods:

• Utilized dynamic neural networks with SAH functions.
• Trained networks to compute SAH for two continuous-variable inputs with varying common mode signals.
• Analyzed hidden unit activity and attractor states after delay periods.

Main Results:

• Networks successfully computed SAH for two inputs, providing insights into associative memory.
• Hidden unit activity in trained networks mimicked primate cortical neuron activity during memory tasks.
• Attractor states were determined by the input-output functions of hidden units, not network architecture.

Conclusions:

• SAH networks offer a viable model for associative short-term memory, particularly for sensory signals with shared components.
• The findings suggest that the properties of individual neurons (hidden units) play a crucial role in memory formation and recall.
• Network architecture is less critical for attractor state determination than the intrinsic properties of its components.