Learning and generation of long-range correlated sequences

Area of Science:

• Computational neuroscience
• Machine learning theory

Background:

• Neural networks are increasingly used for sequence analysis.
• Understanding their ability to capture complex statistical properties is crucial.

Purpose of the Study:

• To investigate if fully connected asymmetric networks can learn and generate power-law correlated sequences.
• To explore how neural networks extract statistical features from data.

Main Methods:

• Training a fully connected asymmetric network on sequences.
• Analyzing the statistical properties of generated sequences.
• Examining the correlation structure within the network's weight matrix.

Main Results:

• The trained network successfully learned the average power-law behavior.
• Generated sequences exhibited similar statistical properties to the training data.
• A correlated weight matrix with power-law characteristics induced similar sequence behavior.

Conclusions:

• Fully connected asymmetric networks are capable of learning and generating power-law correlated sequences.
• The network's ability to extract statistical features is demonstrated.
• Weight matrix correlations play a key role in generating sequences with specific statistical behaviors.