The effect of the non-linear function on system dynamics within delay-feedback reservoirs

Area of Science:

• Computational neuroscience
• Hardware-efficient computing architectures

Background:

• Delay-feedback reservoirs are hardware-efficient computing systems using a single non-linear node and delay line.
• Their versatility stems from the non-linear transform, but optimizing this function for specific tasks is challenging.

Purpose of the Study:

• To explore the impact of different non-linear functions on reservoir dynamics and performance.
• To provide insights into optimizing delay-feedback reservoirs for computational tasks.

Main Methods:

• Compared five non-linear functions: Mackey-Glass, sine squared, double sinusoids, Tan, and Tanh.
• Evaluated performance, system dynamics, and utilization for each function.

Main Results:

• Mackey-Glass showed limited dynamics, excelling in non-linear tasks but failing in memory-intensive ones.
• Sine squared had limited performance; double sinusoid performed well on non-linear tasks.
• Tan exhibited sensitivity similar to Mackey-Glass, while Tanh provided balanced performance across task types.

Conclusions:

• Non-linear function choice significantly impacts delay-feedback reservoir dynamics and task performance.
• Tailoring reservoir dynamics through non-linear function selection is crucial for optimization.