Spontaneous sparse learning for PCM-based memristor neural networks.

Area of Science:

• Neuromorphic Computing
• Materials Science

Background:

• Neural networks excel at intelligent tasks but face training limitations with memristors due to material properties.
• Phase-change memory (PCM) memristors exhibit a resistance drift effect that complicates conventional training methods.

Purpose of the Study:

• To develop a novel training scheme for memristor-based neural networks that utilizes intrinsic device properties.
• To improve the efficiency and performance of neuromorphic computing hardware.

Main Methods:

• A 39nm 1Gb PCM memristor array was fabricated and characterized for its resistance drift.
• A Spontaneous Sparse Learning (SSL) scheme was developed, leveraging resistance drift as a consistency-based distillation process.
• The SSL scheme was applied to handwritten digit classification tasks.

Main Results:

• The SSL scheme demonstrated improved network convergence and performance in handwritten digit classification.
• The method provided enhanced sparsity controllability without requiring additional computational resources.
• The resistance drift effect was successfully harnessed to reinforce array weights.

Conclusions:

• The Spontaneous Sparse Learning scheme effectively integrates memristor's intrinsic properties into neural network training.
• This work advances neuromorphic computing by enabling efficient learning algorithms on memristor devices.
• It opens new avenues for developing next-generation neuromorphic computing chips.