Table 1 Review of the MNIST Literature on neuromorphic hardware

Publication	Hardware	Learning Mode	Network Structure	Energy per Sample (mJ)	Latency per Sample (ms)	Test Accuracy (%)
On-chip backpropagation
Renner et al. (2024) [This study]	Loihi	on-chip nBP	400-400-10^a	0.59	1.5	96.3 (97.5^b)
On-chip single layer training or BP alternatives
⁴⁰ Shrestha et al. (2021)	Loihi	EM-STDP FA/DFA	CNN-CNN-100-10	8.4	20	94.7
³⁹ Frenkel et al. (2020)	SPOON	DRTP	CNN-10	0.000366^c	0.12	95.3
³⁷ Park et al. (2019)	unnamed	mod. SD	784-200-200-10	0.000253^c	0.01	98.1
¹⁰² Chen et al. (2018)	unnamed	S-STDP	236-20^d	0.017	0.16	89
³⁴ Frenkel et al. (2018)	ODIN	SDSP	256-10	0.000015	-	84.5
¹⁰¹ Lin et al. (2018)	Loihi	S-STDP	1920-10^d	0.553	-	96.4
³⁶ Buhler et al. (2017)	unnamed	LCA features	256-10	0.000050	0.001^c	88
On-chip inference only
Renner et al. (2024) [This study]	Loihi	inference	400-400-10^a	0.0025	0.17	96.3 (97.5^b)
⁴⁰ Shrestha et al. (2021)	Loihi	inference	CNN-CNN-100-10	2.47	10	94.7
³⁹ Frenkel et al. (2020)	SPOON	inference	CNN-10	0.000313	0.12	97.5
⁸⁵ Göltz et al. (2019)	BrainScaleS-2	inference	256-246-10	0.0084	0.048	96.9
¹⁰¹ Lin et al. (2018)	Loihi	inference	1920-10^d	0.0128^e	-	96.4
¹⁰² Chen et al. (2018)	unnamed	inference	784-1024-512-10	0.0017	-	97.9
¹⁰⁶ Esser et al. (2015)	True North	inference	CNN (512 neurons)	0.00027	1	92.7
¹⁰⁶ Esser et al. (2015)	True North	inference	CNN (3840 neurons)	0.108	1	99.4
¹⁰⁷ Stromatias et al. (2015)	SpiNNaker	inference	784-500-500-10	3.3	11	95

The table includes three relevant classes of literature. Note that the energy-delay product may be computed from the Energy per Sample and Latency per Sample columns. For a table of simulated SNN learning algorithms, see Supplementary Table IV. Abbreviations: EM-STDP Error-modulated spike-timing dependent plasticity, DFA Direct feedback alignment, DRTP Direct random target projection, SD Segregated dendrites, SDSP Spike-driven synaptic plasticity, LCA Locally competitive algorithm.
^a400 (20 × 20) corresponds to 784 (28 × 28) after cropping the empty (0-padded) image margin of 4 pixels. Including these pixels does not affect accuracy and has a minor effect on inference energy but roughly doubles energy for training, see Supplementary Table II.
^bValue in parentheses is top-1 accuracy read out from last-layer activity (membrane potential) before thresholding.
^cCalculated from given values.
^dOff-chip preprocessing.
^eDynamic energy reported in the Supplementary Material of ⁸².

Search