Improving Recall in Sparse Associative Memories That Use Neurogenesis

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Developing low-power neuromorphic systems necessitates specialized algorithms like spiking neural networks (SNNs).
• Recalling learned patterns from noisy data is a key challenge for SNNs.
• Sparse Associative Memory (SAM) models, inspired by neural coding, and neurogenesis-inspired lifelong learning offer potential solutions but face limitations in capacity and noise resilience.

Purpose of the Study:

• To present a unifying framework for characterizing SAM networks pre-trained with a neurogenesis-inspired learning strategy.
• To formally define network topology and threshold optimization methods for improved performance.
• To enhance memory capacity and noise robustness in neuromorphic associative memory.

Main Methods:

• Developed a unifying framework for characterizing SAM networks.
• Defined network topology and threshold optimization techniques.
• Employed a learning strategy incorporating a neurogenesis model for pre-training.

Main Results:

• Achieved over 10^4 times improvement in memory capacity compared to previous methods.
• Demonstrated reduced interneuron connectivity while maintaining high recall efficacy.
• Validated the effectiveness of the proposed optimization methods.

Conclusions:

• The proposed framework and optimization methods significantly advance SAM networks for neuromorphic applications.
• These advancements pave the way for more efficient and robust low-power associative memory on neuromorphic platforms.
• This research contributes to the development of effective lifelong learning capabilities in artificial systems.