Table 3.

Comparison table of memristive-device-based SNNs for MNIST handwritten recognition.

	This work	Boybat et al., 2018	Querlioz et al., 2011
Training method	Greedy	Conventional	Conventional
Network structure	784 × 50	784 × 50	784 × 50
Accuracy with variations	~75%	~70%	~80%
Devices per synapse	1	≥9	1
Learning increments, decrements	~0.5, ~0.3	0.01, 0.006	0.01, 0.005
Required device levels	~20	~20	>200

The accuracy with variations of this work is obtained with 30% cycle-to-cycle and device-to-device A₊, A₋ variation, and 10% cycle-to-cyle and device-to-device W_max, W_min variation. For Boybat et al. (2018), the N-in-1 architecture (non-differential) with N=9 and with device variation model is listed. And for Querlioz et al. (2011), the data is obtained with 25% cycle-to-cycle A₊, A₋ variation, and 25% cycle-to-cycle W_max, W_min variation.