Table 2 Test accuracy and spiking sparsity for a VGG16 architecture

Dataset	Classes	Test accuracy [%] w/o FT		Test accuracy [%] w/ FT		SNN
		ReLU	SNN	ReLU	SNN	Sparsity
CIFAR10	10	93.59⁶⁷	93.59	93.69 ± 0.02	93.69 ± 0.02	0.38
CIFAR10⁵²	10	–	–	–	91.90	0.24
CIFAR10⁵³	10	–	–	–	92.68	0.62
CIFAR10 + L1	10	92.82	92.82	93.28 ± 0.02	93.27 ± 0.02	0.20
CIFAR100	100	70.48⁶⁷	70.48	72.23 ± 0.06	72.24 ± 0.06	0.38
CIFAR100⁵²	100	–	–	–	65.98	0.28
CIFAR100 + L1	100	69.33	69.33	72.20 ± 0.04	72.21 ± 0.04	0.24
PLACES365	365	52.69⁶⁵	52.69	53.86 ± 0.02	53.86 ± 0.02	0.54
PLACES365 + L1	365	48.67	48.67	48.88 ± 0.06	48.85 ± 0.06	0.27

The first column identifies the dataset and indicates whether L1 regularization was used. The second column gives the number of classes for each task; the third and fourth columns show accuracy after pretraining (ReLU) and conversion (SNN) without fine-tuning (w/o FT); and the fifth and sixth columns show the final results after fine-tuning (w/ FT). The right-most column presents the average number of spikes per neuron (sparsity). The final results obtained in this work are in bold. The standard deviations are reported from 16 trials.

Quick links

Search