Analysis and fully memristor-based reservoir computing for temporal data classification.

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computational Neuroscience

Background:

• Reservoir computing (RC) excels at spatiotemporal signal processing, offering lower training costs than traditional recurrent neural networks.
• Hardware implementation of RC relies on generating dynamic reservoir states.
• Memristors are key components for creating compact and efficient neuromorphic hardware.

Purpose of the Study:

• To develop and evaluate a novel dual-memory reservoir computing system.
• To integrate short-term memory using WOₓ-based memristors and long-term memory using TiOₓ-based memristors.
• To assess the system's performance on temporal data processing tasks.

Main Methods:

• A dual-memory RC system was designed, incorporating WOₓ memristors for short-term memory (16 states, 4 bits) and TiOₓ memristors for long-term memory.
• The memristor characteristics were thoroughly analyzed.
• The RC system was applied to isolated spoken digit recognition and Mackey-Glass time series prediction.

Main Results:

• The system achieved 98.84% accuracy in isolated spoken digit recognition.
• It demonstrated a low normalized root mean square error (NRMSE) of 0.036 in Mackey-Glass time series prediction.
• The WOₓ and TiOₓ memristors effectively contributed to the system's temporal data processing capabilities.

Conclusions:

• Memristor-based RC systems are highly capable of handling complex temporal challenges.
• The proposed dual-memory architecture enhances RC performance for spatiotemporal tasks.
• This research paves the way for future advancements in neuromorphic computing hardware and applications.