Experimental demonstration and tolerancing of a large-scale neural network (165,000 synapses), using phase-change memory as the synaptic weight element

Abstract

Using 2 phase-change memory (PCM) devices per synapse, a 3-layer perceptron network with 164,885 synapses is trained on a subset (5000 examples) of the MNIST database of handwritten digits using a backpropagation variant suitable for NVM+selector crossbar arrays, obtaining a training (generalization) accuracy of 82.2% (82.9%). Using a neural network (NN) simulator matched to the experimental demonstrator, extensive tolerancing is performed with respect to NVM variability, yield, and the stochasticity, linearity and asymmetry of NVM-conductance response.

Trade-off between number of conductance states and variability of conductance change in Pr0.7Ca0.3MnO3-based synapse device

“…However, the PCMO-synapses exhibited asymmetric values of DG which can lead to a degradation in the performance of the S-HNN. 10,25 The asymmetry of DG can be defined as an asymmetry ratio (conductance ratio between potentiation and depression), as shown in the (b).…”

“…[10][11][12][13][15][16][17] Even though each synapse device has its advantages, the synapse device needs a simple two-terminal structure for high density, low-power operation, an analogous conductance change, and reliable characteristics for the implementation of the biological-brain-like S-HNN. 5 In these respects, ReRAM can be a promising candidate.…”

“…A variation in DG can degrade the S-HNN because potentiation and depression can occur randomly during the learning process (information storage process) in the S-HNN. 10,25 For example, if one potentiation pulse and one depression pulse were applied sequentially, the final conductance needs to be similar to the initial conductance. Even though the S-HNN has errortolerant computation, a larger variation in DG can lead to a degradation in the computation of the S-HNN.…”

“…Even though the S-HNN has errortolerant computation, a larger variation in DG can lead to a degradation in the computation of the S-HNN. 10,25 In this respect, the similarity of DG during potentiation and depression is an important requirement for a synapse device. To evaluate the similarity of DG, we defined the asymmetry ratio, which is the conductance ratio between potentiation and depression, as shown in Fig.…”

“…This is also an important synaptic characteristic because a larger number of conductance states lead to higher performance of the S-HNN. 5,10,25 Among the evaluated samples in Fig. 5(b), the Al/PCMO sample exhibited the highest asymmetry ratio, which can lead to a degradation in computation in the S-HNN.…”