Code and results accompanying the Learning from Neural Architecture Search paper accepted at SSCI 2020.
| Feature | MNIST | Fashion | CIFAR-10 |
|---|---|---|---|
| disp.diff_mean_02 | 0.0223 ± 0.0403 | 0.199 ± 0.0736 | 0.0499 ± 0.0459 |
| disp.diff_mean_05 | 0.0223 ± 0.0403 | 0.199 ± 0.0736 | 0.0499 ± 0.0459 |
| disp.ratio_mean_02 | 1 ± 0.00209 | 1.01 ± 0.00381 | 1 ± 0.00238 |
| disp.ratio_mean_05 | 1 ± 0.00209 | 1.01 ± 0.00381 | 1 ± 0.00238 |
| distr.kurtosis | -1.87 ± 0.00648 | -0.718 ± 0.0843 | 4.9 ± 0.258 |
| distr.skewness | -0.0862 ± 0.0289 | -0.87 ± 0.0396 | 2.24 ± 0.0404 |
| ic.eps.max | 0.00323 ± 0.00127 | 0.00462 ± 0.000634 | 0.00247 ± 0.000431 |
| ic.eps.ratio | -1.36 ± 0.021 | -1.73 ± 0.0282 | -2.19 ± 0.0301 |
| ic.eps.s | -1.07 ± 0.00981 | -1.18 ± 0.0134 | -1.31 ± 0.0104 |
| ic.h.max | 0.823 ± 0.0101 | 0.825 ± 0.00952 | 0.768 ± 0.0105 |
| ic.m0 | 0.541 ± 0.0135 | 0.599 ± 0.014 | 0.558 ± 0.00953 |
| lin_simple.adj_r2 | 0.436 ± 0.00781 | 0.383 ± 0.0118 | 0.472 ± 0.0117 |
| lin_simple.intercept | 0.581 ± 0.00636 | 0.617 ± 0.00569 | 0.199 ± 0.00276 |
| lin_w_interact.adj_r2 | 0.327 ± 0.0202 | 0.314 ± 0.0216 | 0.566 ± 0.031 |
| nbc.dist_ratio.coeff_var | 0.132 ± 0.00224 | 0.1 ± 0.00139 | 0.124 ± 0.00206 |
| nbc.nb_fitness.cor | -0.43 ± 0.0119 | -0.461 ± 0.0108 | -0.306 ± 0.00833 |
| nbc.nn_nb.cor | 0.411 ± 0.0213 | 0.589 ± 0.0138 | 0.468 ± 0.0145 |
| nbc.nn_nb.mean_ratio | 0.874 ± 0.00215 | 0.916 ± 0.00104 | 0.889 ± 0.00195 |
| nbc.nn_nb.sd_ratio | 0.539 ± 0.0126 | 0.626 ± 0.0117 | 0.551 ± 0.0124 |
| quad_simple.adj_r2 | 0.486 ± 0.00733 | 0.497 ± 0.0112 | 0.488 ± 0.012 |
Table: ELA features for MNIST, CIFAR-10 and Fashion NAS problem instances.