Generation of unpredictable time series by a neural network

Area of Science:

• Artificial Intelligence
• Complex Systems Dynamics

Background:

• Perceptrons are fundamental units in artificial neural networks.
• Understanding time series generation is crucial for modeling dynamic systems.

Purpose of the Study:

• To investigate a perceptron model that learns its own inverse output.
• To analyze the properties and dynamics of the generated time series.

Main Methods:

• Utilizing a perceptron with an inverse learning rule.
• Analyzing weight vectors, sequence properties (cycle length, probability distributions).
• Investigating system behavior with continuous transfer functions.

Main Results:

• Demonstrated suppression of the autocorrelation function.
• Established connections to the Bernasconi model.
• Observed chaotic and intermittent behavior controlled by learning rate and amplification.

Conclusions:

• The inverse learning perceptron offers a unique method for time series generation.
• The system exhibits complex dynamics, including chaos, under specific conditions.