Table 4 Accuracy per layer and statistical features of their filters for VGG-16 trained on \(K\) labels from CIFAR-100.

Layer	\({N}_{f}\)	\({F}_{s}\)	\(F{C}_{s}\)	Accuracy	\(n\)	\({N}_{c}\)	\({C}_{s}\)
VGG-16 on CIFAR-10/100
13	512	1 × 1	512	0.926	3.26	1.01	2.2
10	512	2 × 2	2048	0.931	4.86	1.83	1.6
7	256	4 × 4	4096	0.908	10.11	1.47	1.7
4	128	8 × 8	8192	0.890	15.83	1.6	1.8
2	64	16 × 16	16,384	0.829	18.64	1.6	2.0
VGG-16 on CIFAR-20/100
13	512	1 × 1	512	0.9115	9.92	1.02	3.7
10	512	2 × 2	2048	0.9115	13.6	2.33	1.9
7	256	4 × 4	4096	0.9065	33.6	1.64	2.31
4	128	8 × 8	8192	0.8465	57	2	2.4
2	64	16 × 16	16,384	0.752	68.23	1.83	2.7
VGG-16 on CIFAR-40/100
13	512	1 × 1	512	0.8553	51.8	1.11	7.5
10	512	2 × 2	2048	0.8567	12.3	2.92	2
7	256	4 × 4	4096	0.7825	38.4	2.44	2.17
4	128	8 × 8	8192	0.6388	143.8	3.22	2.54
2	64	16 × 16	16,384	0.5380	203.6	3.5	2.7
VGG-16 on CIFAR-60/100
13	512	1 × 1	512	0.8277	123.9	1.3	8.13
10	512	2 × 2	2048	0.8275	18.17	2.78	2.3
7	256	4 × 4	4096	0.7148	39.52	2.16	2.24
4	128	8 × 8	8192	0.5392	260.6	4.16	2.6
2	64	16 × 16	16,384	0.4480	423.92	4.5	3

The results are similar to those of Table 1, where VGG-16 was trained on \(K=10, 20, 30,\) and \(60\) labels out of \(100,\) namely CIFAR-K/100 (Supplementary Information).

Quick links

Search