Experimental demonstration of software-trained neural network inferencing in analog memristor crossbar arrays
Abstract: Memristor crossbar structures are known to naturally implement signal accumulation and multiplication, which is particularly relevant in neuromorphic computing. Most work assumes that the physical current accumulation result is close to the ideal mathematical vector matrix multiplications (VMM). Here we address some of the non-idealities, and demonstrate VMM in a 64x64 memristor crossbar array. This work includes a feedbackprogramming scheme to adjust device conductances to target matrix values, and experiments to evaluate performance. We demonstrate the MNIST pattern recognition task using weight matrices taken directly from software training algorithms. We demonstrate that analog crossbar-based neuromorphic circuits can directly utilize existing machine learning algorithms with minimal modifications.