Data Clustering using Memristor Networks.

摘要: Memristors have emerged as a promising candidate for critical applications such non-volatile memory well non-Von Neumann computing architectures based on neuromorphic and machine learning systems. In this study, we demonstrate that memristors can be used to perform principal component analysis (PCA), an important technique data feature learning. The conductance changes of in response voltage pulses are studied modeled with internal state variable trace the analog behavior device. Unsupervised, online is achieved memristor crossbar using Sanger’s rule, derivative Hebb’s obtain components. details weights evolution during training investigated over epochs function parameters. effects device non-uniformity PCA network performance further analyzed. We show memristor-based capable linearly separating distinct classes from sensory high clarification success 97.6% even presence large variations.