Reservoir Computing with Delayed Input for Fast and Easy Optimisation.

Area of Science:

• Machine Learning
• Dynamical Systems
• Computational Neuroscience

Background:

• Reservoir computing (RC) is a machine learning paradigm utilizing the dynamics of recurrent neural networks.
• Its training involves optimizing only the output weights, making it suitable for hardware implementation.
• However, RC typically requires extensive task-specific parameter optimization for optimal performance.

Purpose of the Study:

• To investigate if time-delayed inputs can improve reservoir computing performance without parameter optimization.
• To demonstrate the versatility of a single, unaltered reservoir across diverse time series prediction tasks.
• To reduce the computational expense associated with reservoir computing parameter tuning.

Main Methods:

• Implemented reservoir computing models with the addition of time-delayed input signals.
• Evaluated performance on six distinct time series prediction tasks using an unoptimized reservoir.
• Focused on the impact of incorporating appropriate time-delayed inputs on prediction accuracy.

Main Results:

• Achieved good performance on various time series prediction tasks using an unoptimized reservoir with time-delayed inputs.
• A single, unaltered reservoir demonstrated effectiveness across six different prediction tasks.
• The proposed method significantly reduced computational cost compared to traditional parameter optimization.

Conclusions:

• Time-delayed inputs offer a viable strategy to enhance reservoir computing performance without task-specific parameter optimization.
• This approach is particularly beneficial for hardware implementations where parameter tuning is challenging.
• The findings highlight a computationally efficient method for applying reservoir computing to diverse time-dependent problems.