Table 5 Performance of each encoding method using the final hyper-parameters on the LOFAR dataset with second-order LiF neurons

Encoding Method	Accuracy		AUROC		AUPRC		F1
Delta Exposure	0.887	0.294	0.314	0.007	0.621	0.108	0.397	0.009
Delta Exposure + DN	0.992	0.000	0.312	0.000	0.629	0.011	0.395	0.001
Latency	0.992	0.002	0.346	0.012	0.604	0.021	0.474	0.024
Latency + DN	0.989	0.003	0.335	0.005	0.475	0.007	0.438	0.005
Step-Forward-Direct	0.989	0.006	0.351	0.008	0.497	0.015	0.469	0.012
Step-Forward-Direct + DN	0.992	0.000	0.311	0.001	0.693	0.000	0.393	0.002
ANN	0.748	0.165	0.439	0.053	0.608	0.083	0.087	0.027
ANN + DN	0.862	0.160	0.473	0.056	0.559	0.091	0.072	0.032

Each metric is listed as mean and standard deviation. The best scores are bolded. Ten trials were completed for each encoding method. Abbreviations: DNN divisive normalisation, ANN artificial neural network, AUROC area under the receiver operator curve, and AUPRC area under the precision recall curve. Latency encoding without divisive normalisation offers the best overall performance in accuracy and F1 scores. Step-Forward (direct) encoding with divisive normalisation offers significantly higher AUPRC performance than all other methods. An identically sized ANN does perform best in AUROC, however.

Quick links

Search