Noise tolerance of attractor and feedforward memory models

Area of Science:

• Computational neuroscience
• Neural networks
• Information theory

Background:

• Short-term memory relies on persistent neural activity after stimuli cease.
• Attractor and feedforward networks are key models for neural memory systems.

Purpose of the Study:

• Compare feedforward and attractor network performance in noisy environments.
• Analyze how network nonlinearities and noise affect memory capacity.

Main Methods:

• Utilized Fisher information to quantify memory performance.
• Modeled neural networks with varying architectures and nonlinearities.
• Simulated stimulus amplitude encoding under Gaussian noise.

Main Results:

• Feedforward networks outperform linear attractor networks due to noise removal.
• Attractor networks can surpass feedforward networks with limited neuronal dynamic range or noise reset.
• Optimal attractor networks are forgetful without noise reduction but perfect integrators with strong reset.

Conclusions:

• Network architecture's optimal form depends on nonlinearities and noise handling.
• Trade-offs in memory performance are influenced by network constraints.
• Identified conditions favoring attractor versus feedforward networks for stimulus information storage.