Early Recognition of Skin Malignancy in Images Based on Convolutional Networks by Using Dynamic System Model

Table 3

Table of experiments for different transfer learning configurations, minibatch sizes, and division of the dataset. The three best results have been highlighted in bold, italic, and bold-italic successively.

Experiment	Trained layers (AlexNet)	Minibatch	Learning rate	Better accuracy
				Train	Val
1	All pretrained	16	1.00E-02	7.48E-01	6.69E-01
2		16	1.00E-03	7.73E-01	6.78E-01
3		16	1.00E-04	7.69E-01	6.86E-01
4		32	1.00E-02	7.60E-01	6.64E-01
5		32	1.00E-03	7.79E-01	6.85E-01
6		32	1.00E-04	7.63E-01	6.85E-01
7		64	1.00E-02	7.65E-01	6.80E-01
8		64	1.00E-03	7.76E-01	6.96E-01
9		64	1.00E-04	7.53E-01	6.68E-01
10		128	1.00E-02	7.76E-01	6.83E-01
11		128	1.00E-03	7.72E-01	6.82E-01
12		128	1.00E-04	7.40E-01	6.53E-01
13	Last (FC-8)	16	1.00E-02	7.46E-01	6.49E-01
14		16	1.00E-03	7.76E-01	6.96E-01
15		16	1.00E-04	7.69E-01	6.84E-01
16		32	1.00E-02	7.51E-01	6.64E-01
17		32	1.00E-03	7.81E-01	6.85E-01
18		32	1.00E-04	7.63E-01	6.77E-01
19		64	1.00E-02	7.64E-01	6.66E-01
20		64	1.00E-03	7.79E-01	6.86E-01
21		64	1.00E-04	7.53E-01	6.61E-01
22		128	1.00E-02	7.73E-01	6.85E-01
23		128	1.00E-03	7.72E-01	6.85E-01
24		128	1.00E-04	7.42E-01	6.54E-01