做一个分类mnist 0,2的二分类三层网络,隐藏层节点数由3-1000共实验了59组值。固定收敛标准δ=1e-6,每组值迭代1999次,统计平均分类准确率pave和迭代次数的分布。
统计得到的表格
隐藏层节点数 | 迭代次数的均值 | 平均分类准确率 | δ | 耗时ms/次 | 峰值占比% | 不同峰值数量 | 1/n^2 |
3 | 178616.4877 | 0.986592451 | 1.00E-06 | 796.08804 | 6.8 | 683 | 0.111111 |
5 | 117136.8594 | 0.98855914 | 1.00E-06 | 621.7969 | 2.8 | 674 | 0.04 |
10 | 80329.23862 | 0.988490518 | 1.00E-06 | 583.22361 | 18.4 | 162 | 0.01 |
20 | 56105.15958 | 0.987261772 | 1.00E-06 | 732.94347 | 46.5 | 40 | 0.0025 |
30 | 34663.66183 | 0.985049185 | 1.00E-06 | 606.98649 | 36.7 | 23 | 0.001111 |
40 | 18375.36068 | 0.983432571 | 1.00E-06 | 453.7964 | 37.9 | 13 | 0.000625 |
50 | 12000.17909 | 0.982436049 | 1.00E-06 | 382.18659 | 34.1 | 13 | 0.0004 |
60 | 8716.448224 | 0.982761759 | 1.00E-06 | 335.53677 | 62.5 | 8 | 0.000278 |
70 | 7765.041521 | 0.981953203 | 1.00E-06 | 357.97199 | 82.3 | 6 | 0.000204 |
80 | 7575.328664 | 0.981890548 | 1.00E-06 | 389.42671 | 80.5 | 7 | 0.000156 |
90 | 6987.775888 | 0.982819939 | 1.00E-06 | 417.70035 | 47 | 7 | 0.000123 |
100 | 5530.548274 | 0.98257404 | 1.00E-06 | 423.09355 | 55.3 | 9 | 0.0001 |
110 | 4701.468734 | 0.981943755 | 1.00E-06 | 384.8024 | 54.5 | 9 | 8.26E-05 |
120 | 4529.92096 | 0.981862701 | 1.00E-06 | 392.42271 | 91.8 | 5 | 6.94E-05 |
130 | 4516.147074 | 0.981792089 | 1.00E-06 | 511.30415 | 99.8 | 2 | 5.92E-05 |
150 | 4476.825413 | 0.9805621 | 1.00E-06 | 574.1906 | 95.3 | 4 | 4.44E-05 |
170 | 3694.712356 | 0.954447154 | 1.00E-06 | 565.87844 | 97.3 | 3 | 3.46E-05 |
190 | 3592.71936 | 0.95640887 | 1.00E-06 | 603.2051 | 78.9 | 4 | 2.77E-05 |
210 | 3288.171586 | 0.976363927 | 1.00E-06 | 629.10155 | 98.7 | 4 | 2.27E-05 |
230 | 3272.45923 | 0.975947964 | 1.00E-06 | 631.7939 | 90.6 | 4 | 1.89E-05 |
250 | 3138.144072 | 0.978781637 | 1.00E-06 | 712.0065 | 56.8 | 5 | 0.000016 |
270 | 2888.591296 | 0.973652084 | 1.00E-06 | 773.10255 | 51.4 | 6 | 1.37E-05 |
290 | 2357.346673 | 0.935567436 | 1.00E-06 | 696.62481 | 97.7 | 4 | 1.19E-05 |
310 | 1856.995498 | 0.929918737 | 1.00E-06 | 612.47474 | 62.5 | 2 | 1.04E-05 |
330 | 1556.47924 | 0.92568501 | 1.00E-06 | 606.68534 | 99.4 | 2 | 9.18E-06 |
350 | 1552 | 0.921664858 | 1.00E-06 | 664.22911 | 100 | 1 | 8.16E-06 |
370 | 1552 | 0.918690707 | 1.00E-06 | 772.15358 | 100 | 1 | 7.3E-06 |
390 | 1552 | 0.916166582 | 1.00E-06 | 720.62431 | 100 | 1 | 6.57E-06 |
410 | 1552 | 0.914367472 | 1.00E-06 | 843.29965 | 100 | 1 | 5.95E-06 |
430 | 1551.701351 | 0.912146928 | 1.00E-06 | 793.67234 | 99.8 | 2 | 5.41E-06 |
450 | 1522.433717 | 0.880051358 | 1.00E-06 | 859.56378 | 85.1 | 2 | 4.94E-06 |
470 | 1402.874437 | 0.754977141 | 1.00E-06 | 764.50125 | 74.9 | 2 | 4.53E-06 |
490 | 1347.812906 | 0.722088927 | 1.00E-06 | 886.8004 | 89.8 | 5 | 4.16E-06 |
510 | 1146.158579 | 0.857728317 | 1.00E-06 | 892.62881 | 50.6 | 5 | 3.84E-06 |
530 | 977.2226113 | 0.826342346 | 1.00E-06 | 643.85993 | 97.8 | 3 | 3.56E-06 |
550 | 970 | 0.82386397 | 1.00E-06 | 680.67684 | 100 | 1 | 3.31E-06 |
570 | 970 | 0.822728213 | 1.00E-06 | 702.85343 | 100 | 1 | 3.08E-06 |
590 | 970 | 0.822263766 | 1.00E-06 | 722.5923 | 100 | 1 | 2.87E-06 |
610 | 970 | 0.822133731 | 1.00E-06 | 736.7959 | 100 | 1 | 2.69E-06 |
630 | 970 | 0.822198624 | 1.00E-06 | 735.29915 | 100 | 1 | 2.52E-06 |
650 | 970 | 0.822440544 | 1.00E-06 | 798.82191 | 100 | 1 | 2.37E-06 |
670 | 970 | 0.82361832 | 1.00E-06 | 850.35468 | 100 | 1 | 2.23E-06 |
690 | 970 | 0.824684708 | 1.00E-06 | 854.21961 | 100 | 1 | 2.1E-06 |
710 | 969.873937 | 0.826557165 | 1.00E-06 | 894.21061 | 99.8 | 2 | 1.98E-06 |
730 | 968.3561781 | 0.830087758 | 1.00E-06 | 912.29165 | 98.1 | 3 | 1.88E-06 |
750 | 946.4872436 | 0.851308358 | 1.00E-06 | 929.28664 | 82.9 | 4 | 1.78E-06 |
770 | 809.911956 | 0.929636538 | 1.00E-06 | 851.93297 | 51.7 | 7 | 1.69E-06 |
790 | 716.0025013 | 0.963690593 | 1.00E-06 | 853.78639 | 88.2 | 7 | 1.6E-06 |
810 | 690.8169085 | 0.965884533 | 1.00E-06 | 890.66033 | 82.7 | 5 | 1.52E-06 |
830 | 679.128064 | 0.95942877 | 1.00E-06 | 874.05003 | 75.3 | 5 | 1.45E-06 |
850 | 615.8854427 | 0.869436458 | 1.00E-06 | 898.92997 | 64.6 | 7 | 1.38E-06 |
870 | 525.803902 | 0.763393128 | 1.00E-06 | 798.73787 | 36.9 | 6 | 1.32E-06 |
890 | 331.6188094 | 0.556186393 | 1.00E-06 | 789.82491 | 80.1 | 8 | 1.26E-06 |
910 | 267.3936968 | 0.490041243 | 1.00E-06 | 709.89045 | 99 | 6 | 1.21E-06 |
930 | 264.8129065 | 0.494703117 | 1.00E-06 | 766.3942 | 98.3 | 2 | 1.16E-06 |
950 | 264.1690845 | 0.520551031 | 1.00E-06 | 743.35218 | 92.4 | 2 | 1.11E-06 |
970 | 263.4097049 | 0.550442219 | 1.00E-06 | 760.8024 | 85.5 | 2 | 1.06E-06 |
990 | 261.1205603 | 0.639628711 | 1.00E-06 | 820.14357 | 64.7 | 2 | 1.02E-06 |
1000 | 257.9674837 | 0.762096754 | 1.00E-06 | 814.48074 | 63.9 | 2 | 0.000001 |
Pave
这次实验n=5时pave取得最大pave=0.988559140405195,观察图片pave随n的增加至少有比较明显的两个谷和一个峰,表现了非常不规则的变化,也就是在固定δ的情况下说n越大网络性能越强或者说n越小网络分辨性能越强都是不准确的。
峰值占比
这条曲线一个比较有价值的分割点是n=60,当n>60以后最大峰值占比就几乎都是大于40%,
而峰值占比小于20%的只有3组数据分别是n=3,5,10而这三组数据的平均分辨准确率也是所有59组数据中排名位于3,1,2的前三组数据。这个现象也证实了峰值占比和分辨准确率之间的强烈的正相关关系。
不同峰值数量
这组数据变化幅度非常大暗示和1/n^2之一强烈的关系。不同峰值数量大于100的只有3,5,10这3组数据。
这个现象提供了无需测试集找到网络最优结构的迭代思路:
在固定收敛标准下,在相同的迭代次数内,如果A网络的不同峰值数量>B网络的不同峰值数量,则A网络性能优于B网络。