Memristor-based circuit design of BiLSTM network

Area of Science:

• Artificial Intelligence
• Computer Engineering
• Materials Science

Background:

• Traditional bidirectional long short-term memory (BiLSTM) networks face computational bottlenecks due to extensive parameter calculations.
• The need for faster and more efficient deep learning hardware architectures is increasing.

Purpose of the Study:

• To propose a novel memristor-based bidirectional long short-term memory (MBiLSTM) network.
• To leverage in-memory and parallel computing capabilities of memristors for accelerated AI computations.

Main Methods:

• Designed an MBiLSTM network circuit integrating normalization, memristor-based LSTM, ResNet, dense, and winner-take-all (WTA) circuits.
• Utilized memristor circuits for feature extraction, performance enhancement, output dimension matching, and signal selection.
• Validated the MBiLSTM network through gait recognition and handwritten digit recognition experiments.

Main Results:

• Demonstrated accelerated parameter computation speeds compared to conventional BiLSTM networks.
• Successfully applied the MBiLSTM network to complex tasks like gait and handwritten digit recognition.
• Analyzed the stability, robustness, and manufacturing error tolerance of the memristor-based architecture.

Conclusions:

• The MBiLSTM network offers a promising approach for high-speed AI processing by utilizing memristor crossbar arrays.
• Memristor-based computing can significantly enhance the efficiency of deep learning models.
• The proposed architecture shows potential for real-world applications requiring rapid data analysis and recognition.