Memristive Crossbar Array-Based Probabilistic Graph Modeling.

Area of Science:

• Computational science and engineering
• Materials science and engineering
• Artificial intelligence and machine learning

Background:

• Modern graph datasets exhibit structural complexity and uncertainties, necessitating advanced modeling beyond conventional approaches.
• Existing methods often rely on slow, less reliable sequential operations for probabilistic graph modeling.
• Memristive devices offer unique properties like probabilistic switching and memory, suitable for novel computational paradigms.

Purpose of the Study:

• To introduce a novel memristive crossbar array (CBA)-based probabilistic graph model (C-PGM).
• To address the complexities and uncertainties in structural graph data using memristor characteristics.
• To enable fast processing and massive implementation of probabilistic units for graph analysis.

Main Methods:

• Utilizing Cu$_{0.3}$Te$_{0.7}$/HfO$_{2}$/Pt memristors with probabilistic switching, self-rectifying, and memory properties.
• Leveraging device-to-device variation across multiple memristive CBAs for parallel probabilistic computation.
• Implementing C-PGM for steady-state estimation and the PageRank algorithm on simulated large-scale graphs.

Main Results:

• The hardware-based C-PGM successfully expresses small-scale probabilistic graphs with minimal error in aggregate probability calculations.
• C-PGM enables fast processing and massive implementation of probabilistic units, albeit at the expense of chip area.
• C-PGM-based steady-state estimation and PageRank algorithm achieved comparable accuracy to conventional methods.

Conclusions:

• The memristive crossbar array-based probabilistic graph model (C-PGM) offers a feasible hardware solution for complex graph data.
• C-PGM significantly reduces computational costs for graph analysis tasks like steady-state estimation and PageRank.
• This approach overcomes limitations of sequential processing, paving the way for efficient, large-scale probabilistic graph computation.