Pixel Recurrent Neural Networks

• p. 3 Section 3, opening paragraph

  In Section 3.4 we describe the softmax layer that computes the discrete joint distribution

• p. 4 Section 3.4 heading

  3.4. Masked Convolution

Judge's note: Read Section 3 opening on p3 and Section 3.4 on p4-5. Section 3.4 covers the RGB feature split and masks A and B, with no description of a softmax layer. The softmax is covered in Section 2.2, Table 1 and Section 5.3. The masking half of the pointer is correct, so this is a partly inaccurate roadmap sentence.

Second check: another model agreed

• p. 5 Table 1, Diagonal BiLSTM column

  s-s: 1 × 2 no mask

• p. 4 Section 3.2

  convolution Kss that has a kernel of size 2 × 1

• p. 3 Figure 3 caption

  uses a kernel of size 2 × 1

• p. 3 Figure 3 caption

  uses a kernel of size 2 × 1.

Only the authors can say what is right here.

Judge's note: Table 1 on p5 prints '1 × 2'. Section 3.2 on p4 prints '2 × 1' three times and the Figure 3 caption on p3 prints '2 × 1'. The Row LSTM kernel is 'k × 1' in the text and '3 × 1' in the table, so the table does not simply use a flipped convention throughout. The orientation is not explained; the authors should say which is intended.

Second check: another model agreed

• p. 4 Section 3.2, second paragraph

  results in a map of size n × (2n − 1). At this point we can

• p. 4 Section 3.2, second paragraph

  eration generates a 4h × n × n tensor

Only the authors can say what is right here.

Judge's note: Read on p4. A 1 × 1 convolution keeps spatial size, so applied to the skewed n × (2n − 1) map as the text's order of operations states, it would give 4h × n × (2n − 1). The map returns to n × n only after the later skew-back step. The authors may be counting only valid positions or applying the 1 × 1 convolution before skewing, so this is a query on a dimension statement and does not affect results.

Second check: another model agreed

• p. 7 Section 5.5

  we report the performance of the Diagonal BiLSTM model

• p. 7 Table 4

  PixelCNN: 81.30

• p. 5 Section 4

  For MNIST we use a Diagonal BiLSTM with 7 layers and a value of h = 16

Judge's note: Checked Section 4 on p5 and Section 5.5 with Table 4 on p7. The MNIST PixelCNN and Row LSTM configurations are not specified anywhere in the text. The 1-layer h = 32 variant is described only by its table label. This is an omission in the description, with no numeric conflict.

Second check: another model agreed

• p. 3 Figure 3 caption

  the input map is skewed by offseting each

• p. 3 Figure 3 caption

  offseting each

Judge's note: The misspelling is visible in the Figure 3 caption on the p3 page image.

Second check: none (judge only)

• p. 6 Table 2 caption

  evaluated on the Cifar-10 validation set

Judge's note: The Table 2 caption on p6 prints 'Cifar-10'. The body text, the Table 3 and Table 5 captions and the section headings all use 'CIFAR-10'. This is a capitalisation inconsistency only.

Second check: none (judge only)

• p. 7 Section 5.7, first sentence

  Although to our knowledge the are no published results on

• p. 7 Section 5.7, first sentence

  Although to our knowledge the are no published results

Judge's note: The typo is visible on the p7 page image at the start of Section 5.7.

Second check: none (judge only)

• p. 8 Figure 9 caption, last sentence

  are also inputed relative well (see Figure 1)

• p. 8 Figure 9 caption

  are also inputed relative well

Judge's note: Both wording errors are visible in the Figure 9 caption on the p8 page image. The pointer to Figure 1 (image completions) is appropriate.

Second check: none (judge only)

• p. 8 Section 5.7, second paragraph

  Note that the ImageNet models are in general less com-

• p. 8 Section 5.7, paragraph beginning 'Note that'

  Note that the ImageNet models are in general less com-pressible than the CIFAR-10 images.

Judge's note: Read on the page 8 image. The sentence compares "models" with "images", and the next sentences ("ImageNet has greater variety of images", "The ImageNet images are less blurry") show that images are meant. Wrong word with an evident fix; the meaning is recoverable, so cosmetic.

Second check: none (judge only)

• p. 8 Section 5.7, first paragraph

  straint on model size are currently computation time and

Judge's note: Read on the page 8 image, left column. Subject-verb disagreement as printed; should be "constraints ... are" or "constraint ... is". Meaning is unaffected.

Second check: none (judge only)

• p. 9 References, first entry

  works. pp. 400–406. MIT Press, 2000.

Judge's note: Read on the crop of the page 9 reference list. The entry runs from the paper title straight to "pp. 400–406. MIT Press, 2000" with no venue, whereas the Murray & Salakhutdinov entry in the same list carries "In Advances in Neural Information Processing Systems". The omission is visible on the page; the entry is incomplete rather than wrong. I did not verify what the missing venue is, as that is not in the files.

Second check: none (judge only)

• p. 11 Figure 10 caption

  trained on ImageNet 32x32 (right) images

• p. 11 Figure 10 caption

  Additional samples from a model trained on ImageNet 32x32 (right) images.

• p. 11 Figure 10 caption

  ImageNet 32x32 (right) images

Judge's note: Read on the page 11 image: Figure 10 is one 16x16 grid of samples. The "(right)" matches the Figure 7 caption wording "CIFAR-10 (left) and ImageNet 32x32 (right) images" (text extract, line 629) and looks carried over from it. Stray panel label; delete it.

Second check: none (judge only)