Experimentally validated memristive memory augmented neural network with efficient hashing and similarity search

Area of Science:

• Artificial Intelligence
• Materials Science
• Computer Engineering

Background:

• Lifelong on-device learning requires efficient processing of few samples.
• Current memory-augmented neural networks rely on off-chip memory, hindering practical applications.
• Memristive crossbar platforms offer potential for integrated learning hardware.

Purpose of the Study:

• To experimentally validate the implementation of memory-augmented neural networks on a fully integrated memristive crossbar platform.
• To demonstrate the feasibility of on-chip memory functions for few-shot learning.
• To explore novel functionalities enabled by memristor devices for advanced AI algorithms.

Main Methods:

• Implementation of memory-augmented neural network structures within a memristive crossbar.
• Development of crossbar-based content-addressable memory and locality sensitive hashing.
• Experimental validation and accuracy comparison with digital hardware.
• Simulations for scalability assessment on complex tasks.

Main Results:

• Achieved accuracy closely matching digital hardware for memory-augmented neural networks on memristive crossbars.
• Successfully implemented novel crossbar-based memory and hashing functions.
• Demonstrated scalability for efficient one-shot learning on complex tasks.
• Paved the way for practical on-device lifelong learning.

Conclusions:

• Fully integrated memristive crossbar platforms can effectively host memory-augmented neural networks for on-device lifelong learning.
• Intrinsic memristor properties enable new functionalities for efficient few-shot and attention-based learning.
• This technology advances the development of practical, scalable, and efficient edge AI solutions.