TensorFlow: A System for Large-Scale Machine Learning
arXiv (Cornell University), 2016 · paper · Computer science
Describes TensorFlow, a machine learning system that represents computation as dataflow graphs mapped across machines and devices such as CPUs, GPUs and TPUs. Unlike parameter-server designs, it lets developers experiment with new training algorithms; the authors report performance on real-world applications.
Checked: PDF from usenix.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Substance Figure 1 code: 'layer_1 = tf.nn.relu(tf.matmul(x, W_1) + b_2)' uses b_2, a 10-element bias defined two lines later, where it should use the 100-element bias b_1.
- p. 4 Figure 1, hidden layer line
layer_1 = tf.nn.relu(tf.matmul(x, W_1) + b_2)
Judge's note: The page 4 crop shows b_1 defined as '# 100-element bias vector.' on the line before. The hidden layer then adds b_2, which is defined later as tf.zeros([10]). b_1 is never used anywhere else. As printed, the code uses b_2 before it is defined, and the shapes would not match: a [BATCH,100] result plus a [10] bias.
Second check: another model agreed
Report a problem with this finding
Substance The normalized speedup is defined as 't(b)/t(0) × 50/(50 + b)', but values above 1 (Figure 8(c), axis 1.00–1.10; '9.5%') require t(0)/t(b) × 50/(50+b).
- p. 14 Section 6.3, last paragraph
t(b)/t(0) × 50/(50 + b)
- p. 14 Figure 8(c), right axis
Normalized speedup
Judge's note: Page 14: Figure 8(c) shows the step time falling from about 2.27 s at b=0 to about 1.95 s at b=3. Because t(b) < t(0) and 50/(50+b) ≤ 1, the printed expression is below 1 for every b > 0. It cannot produce the plotted values of 1.00–1.10 or the stated 9.5% gain. The inverted ratio gives about 1.098, consistent with the figure and the text.
Second check: another model agreed
Report a problem with this finding
7 minor (typos and the like; most checked by the judge only)
Wording §4.4 says 'In §6.3 we show how backup workers improve throughput by up to 10%', but §6.3 reports a 9.5% maximum normalized speedup (adjusted for extra resources), not a raw throughput gain.
- p. 12 Section 4.4, end (top of left column)
how backup workers improve throughput by up to 10%
- p. 14 Section 6.3 and Figure 8(c)
adding 3 achieves the highest normalized speedup (9.5%)
Only the authors can say what is right here.
Judge's note: Page 12 has the forward reference. On page 14, §6.3 reports the shortest step time as 1.93 s with 4 backups and the highest normalized speedup as 9.5% with 3 backups. Figure 8(c) shows a step time of about 2.27 s with no backups. The ratio of median step times (2.27/1.93) suggests a raw step-rate gain of about 17%, so 'up to 10%' matches only the resource-normalized figure. The word 'throughput' is ambiguous here. The authors should say whether per-resource throughput was meant.
Second check: another model agreed
Report a problem with this finding
Wording §6.2 says the Sparse curves have 'step times ranging from 5 to 20 ms', but on the same reading of Figure 7 the Sparse points at 100 workers correspond to about 28–30 ms.
- p. 13 §6.2, last paragraph
step times ranging from 5 to 20 ms
- p. 13 Figure 7, Sparse 1GB/16GB curves at 100 workers
Sparse 16GB
Only the authors can say what is right here.
Judge's note: Figure 7's y-axis (batches/s) equals the number of workers divided by the median step time. Every value stated in the text fits this reading. Scalar: 1/1.8 ms is about 555, and 100/8.8 ms is about 11,400. Dense 100MB: 1/0.147 is about 6.8, and 100/0.613 is about 163. Dense 1GB: 1/1.01 is about 1, and 100/7.16 is about 14. All match the plotted points. At 100 workers the Sparse 1GB and 16GB points sit at roughly 3,300–3,600 batches/s, well below the 5,000 that a 20 ms step would give. That implies about 28–30 ms. These values are read from a log-scale plot, so this is a query rather than a definite error.
Second check: another model agreed
Report a problem with this finding
Typo Number agreement error: 'compute gradients for each of the parameter (a backward pass)' should read 'each of the parameters'.
- p. 3 Section 2.1, 'Defining new training algorithms'
dients for each of the parameter (a backward pass)
Judge's note: The crop of page 3, right column, 'Defining new training algorithms', shows the wording exactly as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Stray comma between subject and verb: 'Implementing Momentum in DistBelief [20], required modifications to the parameter server implementation'.
- p. 9 Section 4.1, right column
Momentum in DistBelief [20], required modifications
Judge's note: The page 9 crop, §4.1 right column, shows the comma as quoted. It is punctuation only and does not change the meaning.
Second check: none (judge only)
Report a problem with this finding
Typo Wrong word in §6.2: 'Figure 7 shows that number of null training steps' should read 'shows the number of'.
- p. 13 Section 6.2, first paragraph
Figure 7 shows that number of null train-
Judge's note: The page 13 crop shows 'Figure 7 shows that number of null train-ing steps'.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [37] gives the venue as 'Proceedings of ACL-ICJNLP'; the expected acronym is ACL-IJCNLP.
- p. 18 Reference [37]
ACL-ICJNLP
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the conference acronym is IJCNLP (International Joint Conference on Natural Language Processing). The page 18 crop shows 'ACL-ICJNLP' as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [70] author names are garbled: 'C. tao Chu, S. K. Kim, Y. an Lin' (expected C.-T. Chu, Y.-A. Lin).
- p. 20 Reference [70]
C. tao Chu, S. K. Kim, Y. an Lin
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the authors' given names are Cheng-Tao Chu and Yi-An Lin. The page 20 crop shows 'C. tao Chu' and 'Y. an Lin' as typeset, which looks like a BibTeX mishandling of hyphenated given names.
Second check: none (judge only)