Table 1 Comparison of training methods and network architectures across different datasets

Dataset	Category	Training method	Network architecture	Testing accuracy
N-MNIST²⁶	Ours	BPTT	RLIF	98.33 ± 0.04%
		pp-prop	RLIF	98.25 ± 0.03%
		BPTT	RadLIF	98.29 ± 0.02%
		pp-prop	RadLIF	98.40 ± 0.03%
	Online	ETLP³¹	RLIF	94.30%
		e-prop³¹	RLIF	97.90%
		OSTL⁹	sSNU	96.80 ± 0.17%
SHD²⁵	Ours	BPTT	RLIF	94.01 ± 0.12%
		pp-prop	RLIF	93.93 ± 0.28%
		BPTT	RadLIF	94.72 ± 1.06%
		pp-prop	RadLIF	95.33 ± 0.11%
	Online	ETLP³¹	RLIF	78.71 ± 1.49%
		e-prop³¹	RLIF	80.79 ± 0.39%
		OTPE¹¹	LIF	76.70 ± 0.70%
		e-prop⁶⁰	TC-RLIF	80.57%
		OSTTP³²	SNU	77.33 ± 0.8%
		S-TLLR³³	RLIF	78.24 ± 1.84%
	Offline	EventProp⁶¹	RLIF	93.50 ± 0.70%
		BPTT³⁵	DCLS-Delays	95.07 ± 0.24%
		BPTT²⁹	RadLIF	94.62%
Gesture²⁴	Ours	BPTT	RLIF	93.97 ± 0.15%
		pp-prop	RLIF	94.26 ± 0.32%
		BPTT	EGRU	97.45 ± 0.27%
		pp-prop	EGRU	97.29 ± 0.16%
	Online	FPTT³⁴	RLIF	92.13 ± 0.87%
		FPTT³⁴	CNN	97.22%
		OTTT¹⁰	VGG-11	96.88%
	Offline	BPTT⁶²	STS-ResNet	96.7%
		BPTT⁶³	STL-SNN	97.01 ± 0.23%
		BPTT³⁷	PLIF	97.57%
		BPTT³⁶	VGGSNN	97.57%

Results show testing accuracy with standard deviation for various training approaches, including BPTT and pp-prop applied to different network architectures (RLIF, RadLIF) on N-MNIST, SHD, and Gesture datasets. State-of-the-art online and offline learning methods are included for comparison. The bolded data signifies the optimal performance.

Quick links

Search