Table 1.

Testing accuracy percentage over different datasets and training methods.

Initialisation	Training	N-MNIST	F-MNIST	DVS128	SHD	SSC
Homog.	Standard	97.4 ± 0.0	80.1 ± 7.4	76.9 ± 0.8	71.7 ± 1.0	49.7 ± 0.4
Heterog.	Standard	97.5 ± 0.0	87.9 ± 0.1	79.5 ± 1.0	73.7 ± 1.1	53.7 ± 0.7
Homog	Heterog.	96.6 ± 0.2	79.7 ± 7.4	81.2 ± 0.8	82.7 ± 0.8	56.6 ± 0.7
Heterog.	Heterog.	97.3 ± 0.1	87.5 ± 0.1	82.1 ± 0.8	81.7 ± 0.8	60.1 ± 0.7
Chance level		10.0	10.0	10.0	5.0	2.9

Effect of initialisation and training configuration on performance, on datasets of increasing temporal complexity. Initialisation can be homogeneous (all time constants the same) or heterogeneous (random initialisation), and training can be standard (only synaptic weights learned) or heterogeneous (time constants can also be learned). N-MNIST and F-MNIST are static image datasets with little temporal structure, DVS128 is video gestures, and SHD and SSC are temporally complex auditory datasets.