Reducing the Dimensionality of Data with Neural Networks
Science, 2006 · paper · AI and machine learning
Describes a way of initializing the weights of deep autoencoder networks, which convert high-dimensional data into low-dimensional codes. The authors report that the resulting codes work much better than principal components analysis for reducing dimensionality.
Checked: PDF from cs.toronto.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Reference or label Reference 6 is printed as "G. E. Hinton, Neural Comput. 14, 1711 (2002)."; the first page is probably 1771, not 1711.
- p. 4 References and Notes, ref. 6
G. E. Hinton, Neural Comput. 14, 1711 (2002).
- p. 3 body text after Eq. 2, citing (6)
probability of the training data (6)
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that Hinton's "Training products of experts by minimizing contrastive divergence", Neural Computation 14 (2002), begins on page 1771 (pp. 1771-1800). The page 4 crop clearly shows "14, 1711 (2002)", so this is not an extraction artefact. The citing sentence on page 3 ("not exactly following the gradient of the log probability of the training data (6)") fits the contrastive-divergence paper. The true page number cannot be checked from the files, so the authors or publisher should confirm it.
Second check: another model agreed
Report a problem with this finding
Reference or label Note 13 reads "The Reuter Corpus Volume 2 is available at http://trec.nist.gov/data/reuters/reuters.html."; "Reuter" should be "Reuters", and the 804,414-story collection is probably Reuters Corpus Volume 1 (RCV1-v2), not Volume 2.
- p. 4 References and Notes, ref. 13
The Reuter Corpus Volume 2 is available at
- p. 4 body, middle column, document retrieval paragraph
804,414 newswire stories (13)
- p. 4 References and Notes, entry 13
The Reuter Corpus Volume 2 is available at http://
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the 804,414-document collection is RCV1-v2 (Reuters Corpus Volume 1, version 2) and that Reuters Corpus Volume 2 is a different, multilingual collection. The page 4 crops show "The Reuter Corpus Volume 2" in note 13 and "804,414 newswire stories (13)" in the body. The misspelling "Reuter" can be checked inside the paper: the URL in the same note spells it "reuters". The volume point cannot be checked from the files; the authors may have meant "Volume 2" as shorthand for RCV1-v2.
Second check: another model agreed
Report a problem with this finding
4 minor (typos and the like; most checked by the judge only)
Reference or label Reference 16 is printed as "J. A. Tenenbaum, V. J. de Silva, J. C. Langford, Science 290, 2319 (2000)."; the author initials are probably wrong (J. B. Tenenbaum, V. de Silva).
- p. 4 References and Notes, ref. 16
J. A. Tenenbaum, V. J. de Silva, J. C. Langford, Science
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the authors of the Isomap paper at Science 290, 2319 (2000) are Joshua B. Tenenbaum, Vin de Silva and John C. Langford. The page 4 crop shows the initials "J. A." and "V. J." exactly as quoted, so this is not an extraction artefact. The journal, volume, page and year are consistent with the intended paper, so the reference can still be found. The initials cannot be checked from the files.
Second check: another model agreed · checked by hand
Report a problem with this finding
Typo The body text says "the model is "unfolded" (Fig. 1)", while Fig. 1 labels the step "Unrolling" and its caption says the RBMs are "unrolled".
- p. 4 body, left column, paragraph starting 'After pretraining multiple layers'
the model is “unfolded” (Fig. 1)
- p. 2 Fig. 1 caption and panel label
‘‘unrolled’’ to create a deep autoencoder
- p. 4 body text, paragraph beginning 'After pretraining multiple layers'
the model is ‘‘unfolded’’ (Fig. 1)
Judge's note: On the page 4 image, the left column reads: After pretraining multiple layers of feature detectors, the model is "unfolded" (Fig. 1). The text layer has "Bunfolded[" because the quotation-mark glyphs were mis-extracted, which is why the quote was not located. The page 2 crop shows the panel label "Unrolling" and the caption's "unrolled". The two quoted terms are synonyms for the same step and the meaning is clear, so this is a terminology inconsistency only; either form is legitimate.
Second check: none (judge only)
Report a problem with this finding
Typo Subject-verb disagreement: "The first layer of feature detectors then become the visible units for learning the next RBM"; the singular subject "layer" takes "becomes".
- p. 3 body, right column, last paragraph
layer of feature detectors then become the
Judge's note: Read the crop of page 3, right column. The sentence is printed as quoted. The head noun is the singular "layer", so "become" is a grammatical slip; notional agreement with "feature detectors" is possible, and the meaning is unaffected. Cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo The URL in note 12 is printed as "www.cs.toronto.edu/ roweis/data.html", with a blank where the tilde should be (intended: www.cs.toronto.edu/~roweis/data.html).
- p. 4 References and Notes, ref. 12
cs.toronto.edu/ roweis/data.html.
Judge's note: The page 4 crop shows a clear gap after "cs.toronto.edu/" and no tilde glyph; the text layer has a space at the same place. I cannot rule out that a tilde glyph exists in the PDF and was dropped when the page was rendered, but the image and the text layer agree. The URL as printed is not valid, and the fix is evident.
Second check: none (judge only)