Model-size reduction for reservoir computing by concatenating internal states through time

Area of Science:

• Machine Learning
• Computational Neuroscience
• Edge Computing

Background:

• Reservoir computing (RC) is a powerful machine learning approach for rapid time-series analysis.
• Implementing RC in edge computing necessitates reducing its substantial computational resource demands.
• Existing RC models often require large, high-dimensional dynamical systems (reservoirs).

Purpose of the Study:

• To propose and analyze methods for reducing the size of reservoir computing models.
• To enable efficient RC implementation in resource-constrained edge computing environments.
• To investigate the underlying mechanisms of model-size reduction in RC.

Main Methods:

• Developed novel RC techniques by incorporating past or drifting reservoir states into the output layer.
• Analyzed the model-size reduction mechanism using information processing capacity metrics.
• Evaluated method effectiveness on benchmark time-series prediction tasks (generalized Hénon-map, NARMA).

Main Results:

• Achieved significant reduction in reservoir size, up to one-tenth of the original size.
• Maintained high accuracy in time-series prediction tasks with minimal increase in regression error.
• Demonstrated the efficacy of proposed methods for resource-efficient RC.

Conclusions:

• The proposed methods effectively reduce the computational footprint of reservoir computing.
• This advancement is crucial for deploying sophisticated RC models in edge computing.
• Future work can explore further optimizations for RC in real-world edge applications.