Table 12.

This table shows the conclusions of the experiments conducted in this paper.

No.	Algorithm	Datasets	Experiment	Conclusion
1.	ANN–CNN described in Table 2	N-MNIST, N-Caltech101, DvsGesture	Spikes are summed over the presentation duration and collapsed into images. Then they are trained using an ANN.	The ANN obtains comparable to state-of-the-art results on N-MNIST (99.23%) and N-Caltech101 (78.01%). However, ANN performs significantly worse than state-of-the-art on DvsGesture (71.01%), as it cannot handle the spatiotemporal information. Hence, we conclude that N-MNIST and N-Caltech101 does not have additional information contained in the timing of spikes necessary to classify the dataset.
2.	RD-STDP and STDP-tempotron	N-MNIST and DvsGesture	Comparison of rate-based and temporal STDP algorithms on the two datasets	While the rate-dependent RD-STDP obtains very good performance on N-MNIST (83.89%), it is unable to do as well in DvsGesture (76.13%). In contrast, STDP-tempotron performs better in DvsGesture (59.11%), but worse for N-MNIST (53.18%). We conclude that while DvsGesture has spatio-temporal information, and therefore needs STDP-tempotron, N-MNIST does not have additional information in the time domain necessary to classify it.
3.	RD-STDP	N-MNIST	Fixing the output spike time	Despite fixing the output spike time, the system performs well (84.10%), demonstrating that precise timing of spikes are not useful at all in N-MNIST
4.	Populate rate dependent plasticity (new rule)	N-MNIST	A new STDP curve was devised based on summing up spikes over the population—the instantaneous population firing rate (The second definition of firing rate—see section 1, Paragraph 2)	Despite using an STDP curve based on the population spike rates alone, the system is able to give good performance on N-MNIST (85.45%). This demonstrates that spike timing is not important in classifying N-MNIST.