Adaptive spatial hashing with dual-domain memristive hardware

Area of Science:

• Computer Science
• Electrical Engineering
• Materials Science

Background:

• Locality-sensitive hashing (LSH) is crucial for approximate similarity search but faces hardware limitations due to fixed thresholds and inefficient analog encoding.
• Existing LSH hardware implementations struggle with adaptability and energy efficiency, hindering their deployment in resource-constrained environments.

Purpose of the Study:

• To introduce a novel Dual-Domain Adaptive Spatial Hashing (DASH) architecture for efficient and hardware-native similarity search.
• To overcome the limitations of traditional LSH by integrating analog and digital processing with adaptive capabilities.
• To demonstrate the effectiveness of DASH in improving accuracy, semantic preservation, and noise resilience for similarity search.

Main Methods:

• Developed a DASH architecture on a monolithic one-transistor-one-resistor active array utilizing a multifunctional memristor.
• Implemented dual-domain processing: entropy-maximized random projection and data-driven bias adaptation in the analog domain, followed by Hamming-distance computation in the digital domain.
• Utilized dual-domain vector-matrix multiplication for compressing multidimensional inputs into binary hash codes.

Main Results:

• Experimental validation on synthetic data confirmed DASH's ability to maintain spatial separability and enhance bit entropy through adaptation.
• Large-scale simulations on a digit dataset showed improved semantic preservation, similarity recall, and noise resilience compared to non-adaptive hashing methods.
• The unified memristive hardware platform enabled compact similarity encoding and efficient processing.

Conclusions:

• DASH offers a scalable, hardware-native solution for energy-efficient, locality-aware similarity search.
• The adaptive dual-domain approach significantly improves performance metrics over traditional LSH methods.
• DASH is well-suited for deployment in edge and neuromorphic systems requiring high-performance similarity search capabilities.