Who is the Winner? Memristive-CMOS Hybrid Modules: CNN-LSTM Versus HTM.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Neuromorphic Engineering

Background:

• Biological neural networks exhibit hierarchical, modular, and sparse information processing.
• Artificial neural networks, including Hierarchical Temporal Memory (HTM), aim to replicate these brain-like characteristics.
• Implementing these complex architectures efficiently remains a challenge.

Purpose of the Study:

• To evaluate Convolutional Neural Network (CNN) combined with Long Short-Term Memory (LSTM) as an alternative to HTM for hierarchical, modular, and sparse information processing.
• To compare the performance of CNN-LSTM and HTM on a face recognition task with limited training data.
• To present analog CMOS-memristor circuit implementations for the CNN-LSTM architecture.

Main Methods:

• A comparative performance analysis between CNN-LSTM and HTM was conducted using a face recognition dataset.
• Analog CMOS-memristor circuit blocks were designed and presented for implementing the CNN-LSTM architecture.
• Memristor variability and failure analyses were incorporated into the study.

Main Results:

• The CNN-LSTM architecture, particularly its memristive implementation, demonstrated superior recognition performance compared to HTM.
• The memristor-based CNN-LSTM circuits were found to be easier to implement and train.
• The study successfully presented practical circuit implementations for brain-inspired computing.

Conclusions:

• CNN-LSTM presents a promising alternative for creating hierarchical, modular, and sparse artificial neural networks.
• Memristor technology enables efficient and high-performance hardware implementations of advanced neural network architectures.
• The findings suggest a pathway towards more practical and effective neuromorphic computing systems.