Author-maintained record

Post-publication corrections

Corrections identified after publication of “Spatial frequency loss for mitigating spectral bias in deep neural networks.”

Status. This is an author-maintained correction record. A formal correction request to the publisher is planned. When a publisher-issued correction becomes available, its permanent link will be added here.

1. Table 2, ℒB row

The published ℒB row contains subband-loss values from the seed-20 baseline run. The formally selected baseline run was seed 10, selected because it had the smallest final pixel-wise L2 loss among seeds 10, 20, and 30. Supplementary Table S9 records final PL2L values of 2.713, 2.931, and 3.009 × 10−3, respectively.

The published row is:

ConditionSB1SB2SB3SB4SB5SB6SB7
ℒB, published (seed 20)17.3723.7726.6817.007.5244.78217.89

The corrected row, using the formally selected seed-10 run, is:

ConditionSB1SB2SB3SB4SB5SB6SB7
ℒB, corrected (seed 10)17.1722.8423.0612.406.7475.16724.36

The correction changes only the numerical values in the Table 2 ℒB row. It does not change the caption, main text, conclusions, or the reported effect of SFL.

2. Vertical-axis exponents in Figure 4 and Supplementary Figure S3

LocationPublishedCorrect
Fig. 4(b), vertical axis×10−5×10−4
Fig. 4(d), vertical axis×10−5×10−4
Supplementary Fig. S3(b), vertical axis×10−5×10−4
Supplementary Fig. S3(d), vertical axis×10−5×10−4

Figure 4(a) and 4(c), and Supplementary Fig. S3(a) and S3(c), require no correction. The plotted data, interpretation, text, and conclusions are unchanged.

Supporting data and plotting scripts

The learning-curve data supporting the Table 2 correction and the gnuplot scripts used to reproduce the corresponding learning-curve figures are provided below. Both data files contain 75 rows for epochs 1 through 75. At epoch 75, columns 8 through 14 of the seed-20 data reproduce the published Table 2 values, whereas the same columns of the formally selected seed-10 data reproduce the corrected values shown above.

Learning-curve file format

Each learning-curve file is a whitespace-delimited text file with 15 columns. The Table 2 subband-loss values are taken from columns 8 through 14 at epoch 75.

ColumnContent
1Epoch
2Consistency loss
3Pixel-wise L2 reconstruction loss (PL2L)
4Quantizer loss
5Codebook loss
6Commitment loss
7Weighted subband loss
8Unweighted subband loss, SB1
9Unweighted subband loss, SB2
10Unweighted subband loss, SB3
11Unweighted subband loss, SB4
12Unweighted subband loss, SB5
13Unweighted subband loss, SB6
14Unweighted subband loss, SB7
15Feature reconstruction loss

Although the final column reports the feature reconstruction loss, its weight was set to 0 throughout training. The value was logged only for monitoring and was not included in the training objective.

Reproducing the figures

Run each script from docs/corrigendum/scripts. The scripts read the corresponding file from ../data and write the L2-loss and subband-loss EPS figures to ../figures. The corrected vertical-axis exponent, ×10−4, is used for both subband-loss figures.

Provenance

The Table 2 correction is supported by the selected-run criterion reported in Supplementary Table S9 and by the publicly provided seed-10 and seed-20 learning-curve records. The seed-10 run is the formally selected baseline run. The seed-20 record is included to make the origin of the published Table 2 values directly verifiable.