HOLISMOKES - XI. Evaluation of supervised neural networks for strong-lens searches in ground-based imaging surveys

Open Access

Table 2

Performance of various CNN and ResNet architectures.

Architecture	AUROC	TPR₀	TPR₁₀
Baseline CNN	0.9557	0.0	44.4
CNN v2	0.9115	10.6	21.2
CNN v3	0.9694	0.0	45.0
CNN v4	0.9242	0.0	13.2
CNN v5	0.9772	1.6	38.1
CNN v6	0.9927	10.1	46.0
CNN v7	0.9704	7.4	31.8
G-CNN v1	0.9705	13.8	51.9
G-CNN v2	0.9865	7.9	41.8
G-CNN v3	0.9883	15.9	48.2

Baseline ResNet	0.9913	36.0	55.0
ResNet v2	0.8939	21.2	34.4
ResNet v3	0.9472	28.0	37.0
ResNet v4	0.9499	10.1	22.8
ResNet v5	0.9857	16.9	34.4
ResNet v6	0.9885	12.2	30.2
ResNet v7	0.9884	19.6	43.9
ResNet v8	0.9199	15.9	31.8

Notes. Architectures correspond to variations of the baseline CNN, G-CNN and ResNet introduced in Section 4. Further details are given in Section 5.3. The training set was kept fixed to the baseline.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.