Academic Fraud Investigation Report: Deep EM Capsule Network for Fully Overlapping Handwritten Digit Recognition and Separation

Academic fraud report · Geng Detector

Summary

This report examines a 2020 paper published in Acta Automatica Sinica (DOI: 10.16383/j.aas.c190849) by Yao Hongge et al. on a deep EM capsule network for overlapping handwritten digit recognition. The verdict is highly suspicious. The primary finding concerns Tables 3 and 4, which report per-epoch training times with standard deviations that are uniformly exactly ±2 across every experimental condition (e.g., 150±2, 210±2, 240±2, 300±2, 340±2; 350±2, 410±2, 440±2). Such identical variability across all comparisons is statistically implausible for GPU-based deep learning experiments, which typically exhibit I/O-, thermal-, and scheduling-related fluctuations. This strongly suggests fabricated or back-calculated data rather than genuine training logs. A secondary concern is Table 2, where the proposed DCN's activation vector norms at R=3 (0.9800, 0.9943, 0.9923) across three different datasets appear suspiciously close to unity, raising cherry-picking or fabrication concerns. Table 8 additionally shows unprofessional reporting with blank classification entries. Pixel-level image analysis was not possible due to unavailable figures. Confidence is high for the timing-data anomaly, moderate for the magnitude-data anomaly.

Verdict

Highly suspicious. Multiple data tables exhibit patterns inconsistent with genuine experimental measurement. The strongest red flag is the uniform ±2 standard deviation across all reported epoch timings, which is statistically implausible and strongly indicative of fabricated or back-calculated data.

Key findings

Identical standard deviations in epoch timing data (Tables 3 and 4): Every reported per-epoch time in Tables 3 and 4 carries a standard deviation of exactly ±2 (e.g., 150±2, 210±2, 240±2, 300±2, 340±2 in Table 3; 350±2, 410±2, 440±2 in Table 4). Genuine GPU training times fluctuate due to I/O, thermal throttling, and background processes; identical variability across all conditions is implausible.
Suspiciously uniform activation norms (Table 2): At R=3, the proposed DCN's activation vector norms are 0.9800 (MNIST), 0.9943 (fully overlapping), and 0.9923 (mixed) — all extremely close to 1.0 across three distinct data distributions, suggesting possible cherry-picking or fabrication.
Unprofessional classification reporting (Table 8): Failed recognition cases (e.g., "王, 丑" and "也, 卫") are reported as blank entries with "cannot determine," which is incompatible with softmax or squashing outputs that always produce a maximum-probability node.
Image-based checks not performed: First-style (image reuse) and third-style (image splicing) analyses could not be conducted because original high-resolution figures were not available in the extracted text.

Evidence highlights

Tables 3 and 4 (Page 10) — uniform ±2 SDs across all entries.
Table 3 (DCN vs. CapsNet, R=1, 2, 3): 150±2, 210±2, 240±2, 300±2, 340±2.
Table 4 (iterative routing vs. EM): 350±2, 410±2, 440±2, and other entries, all ±2.
Table 2 (Page 9) — DCN R=3 norms: 0.9800 / 0.9943 / 0.9923 across three datasets.
Table 8 (Page 13) — blank classification cells for failure cases with "cannot determine" annotation.
DOI: 10.16383/j.aas.c190849; received 2019-12-18; online first 2020-06-08.

Notes

The strongest evidence is the universal ±2 standard deviation in Tables 3 and 4; verification of original training logs (per-epoch wall-clock timestamps) would be dispositive.
The activation-norm anomaly in Table 2 is suggestive but not conclusive on its own; it should be evaluated jointly with the timing evidence.
The Table 8 reporting issue is unprofessional and may reflect manual post-hoc judgement rather than network output, but it is not by itself evidence of fabrication.
Pixel-level image reuse and splicing analyses were not feasible in this review and should be attempted with the original PDF figures.
The authors should be asked to provide raw training logs, code, and reproducibility scripts; the journal editorial office should verify the integrity of the experimental record.
All conclusions are provisional pending official institutional investigation. False positives are possible.