CoChat
Sign Up
Cited by 61,544Handwritten Text Recognition TechniquesVehicle License Plate RecognitionNeural Networks and Applications

The DOI matches its Crossref record; 131 of the 131 references it cites were found in a registry or an index; 1 of the 78 sources its claims cite were checked for support.

Gradient-based learning applied to document recognition

Open access
Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.·1998·Proceedings of the IEEE
DOI 10.1109/5.726791
Open this paper in CoChat — free accountOpen PDF

Abstract

Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day.

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

The paper argues that gradient-based learning, especially convolutional neural networks and graph transformer networks, can replace much hand-crafted feature extraction and heuristic module tuning in document recognition. It reports strong handwritten digit results and demonstrates globally trained systems for online handwriting and commercial check reading.

Key findings

  • On MNIST, LeNet-5 reached 0.95% test error on the regular database and 0.8% when trained with artificially distorted examples; boosted LeNet-4 achieved the best raw error at 0.7%.
  • Convolutional networks outperformed several alternatives on the same digit benchmark, including linear classifiers (12%), pairwise linear classifiers (7.6%), K-NN (5.0%), PCA plus polynomial classifier (3.3%), RBF network (3.6%), one-hidden-layer multilayer networks (4.5–4.7%), two-hidden-layer networks (2.95–3.05%), LeNet-1 (1.7%), LeNet-4 (1.1%), and tangent distance (1.1%).
  • Global word-level training improved online handwriting recognition: for HOS with a 25,461-word dictionary, word/character errors fell from 4.6%/2.0% to 3.2%/1.4%; for SDNN/HMM without dictionary constraints, they fell from 38%/12.4% to 26%/8.2%.
  • A GTN-based check reader on 646 machine-printed business checks achieved 82% correct, 1% error, and 17% rejects, compared with 68% correct, 1% error, and 31% rejects for the previous system; it was deployed commercially and read millions of checks per day.

Methods

  • The digit comparison used the MNIST database, built from NIST SD-1 and SD-3, with 60,000 training patterns and a 10,000-pattern test subset (5,000 from each source), size-normalized and centered; LeNet-5 is a seven-layer convolutional network with convolutional, subsampling, fully connected, and RBF output layers.
  • The paper introduces graph transformer networks, where modules take and produce directed graphs whose arcs carry numerical information; training uses backpropagation through differentiable graph transformations, including Viterbi, forward, and discriminative forward criteria.
  • Applications include online handwriting recognition with HOS and SDNN/HMM architectures, and a check amount reading cascade combining field location, segmentation, LeNet-5 character recognition, grammar composition, and Viterbi selection.

Limitations

  • Stated: the paper notes that the absence of harmful local minima in multilayer networks remains theoretically mysterious; that more training data would likely improve LeNet-5 accuracy; that SDNN had not yet outperformed HOS; and that global training had only a minor effect in the check-reading task because only a small subset of parameters was trained globally.
  • Evident from the text: the main quantitative benchmark is limited to isolated handwritten digits, so the reported comparisons may not transfer directly to all document recognition tasks.
  • Evident from the text: the check-reading results come from a specific 646-check business-check test set and a commercial deployment, with limited detail for independent reproduction.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/5.726791

Cites

131 of 131 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 of 78 cited sources checked · 0 supported · 0 flagged · accuracy not yet measured

  • It may state two separate claims as one. 119
  • Probably not a substantial claim. 154
  • The quote only partly states this claim. 6
242 of 242

  • The paper relies on this without support.

    Real-life document recognition systems are composed of multiple modules including field extraction, segmentation, recognition, and language modeling.

    “Real-life document recognition systems are composed of multiple modules including field extraction, segmentation, recognition, and language modeling.”

  • The paper relies on this without support.

    The variability and richness of natural data make it almost impossible to build an accurate recognition system entirely by hand.

    “Since the early days of pattern recognition it has been known that the variability and richness of natural data, be it speech, glyphs, or other types of patterns, make it almost impossible to build an accurate recognition system entirely by hand.”

  • The paper relies on this without support.

    Most pattern recognition systems are built using a combination of automatic learning techniques and hand-crafted algorithms.

    “Consequently, most pattern recognition systems are built using a combination of automatic learning techniques and hand-crafted algorithms.”

  • The paper relies on this without support.

    The feature extractor contains most of the prior knowledge and is rather specific to the task.

    “The feature extractor contains most of the prior knowledge and is rather specific to the task.”

  • The paper relies on this without support.

    The classifier is often general purpose and trainable.

    “The classifier, on the other hand, is often general purpose and trainable.”

  • The paper relies on this without support.

    The recognition accuracy is largely determined by the ability of the designer to come up with an appropriate set of features.

    “One of the main problems with this approach is that the recognition accuracy is largely determined by the ability of the designer to come up with an appropriate set of features.”

  • The paper relies on this without support.

    A large proportion of modern commercial OCR systems use some form of multilayer NN trained with back propagation.

    “As evidence of this fact, a large proportion of modern commercial OCR systems use some form of multilayer NN trained with back propagation.”

  • The paper relies on this without support.

    The back-propagation procedure applied to multilayer NN's with sigmoidal units can solve complicated learning tasks.

    “The third event was the demonstration that the back-propagation procedure applied to multilayer NN's with sigmoidal units can solve complicated learning tasks.”

  • The paper relies on this without support.

    Most practical pattern recognition systems are composed of multiple modules.

    “As stated earlier, most practical pattern recognition systems are composed of multiple modules.”

  • The paper relies on this without support.

    Typically, each module is manually optimized, or sometimes trained, outside of its context.

    “Typically, each module is manually optimized, or sometimes trained, outside of its context.”

  • The paper relies on this without support.

    The input variables can be presented in any (fixed) order without affecting the outcome of the training.

    “The input variables can be presented in any (fixed) order without affecting the outcome of the training.”

  • The paper relies on this without support.

    Images have a strong 2-D local structure: variables that are spatially or temporally nearby are highly correlated.

    “On the contrary, images (or time-frequency representations of speech) have a strong 2-D local structure: variables (or pixels) that are spatially or temporally nearby are highly correlated.”

  • The paper relies on this without support.

    Local correlations are the reasons for the well-known advantages of extracting and combining local features before recognizing spatial or temporal objects.

    “Local correlations are the reasons for the well-known advantages of extracting and combining local features before recognizing spatial or temporal objects, because configurations of neighboring variables can be classified into a small number of categories (e.g., edges, corners, etc.).”

  • The paper relies on this without support.

    Drawing sensible conclusions from learning experiments requires that the result be independent of the choice of training set and test among the complete set of samples.

    “Drawing sensible conclusions from learning experiments requires that the result be independent of the choice of training set and test among the complete set of samples.”

  • The paper relies on this without support.

    Saturation of the sigmoids leads to slow convergence and ill-conditioning of the loss function.

    “Saturation of the sigmoids must be avoided because it is known to lead to slow convergence and ill-conditioning of the loss function.”

  • The authors state this as an assumption.

    The cut generation heuristic will include the correct set of cuts.

    “The cut generation heuristic is designed so as to generate more cuts than necessary in the hope that the”

  • The authors state this as an assumption.

    Penalties can be interpreted as negative log posteriors, so that posterior probabilities for an interpretation combine by summing and penalties combine as the negative logarithm of the sum of negative exponentials.

    “We use a combination rule borrowed from a probabilistic interpretation of the penalties as negative log posteriors.”

  • The authors state this as an assumption.

    A particular label sequence is given as the desired sequence to the GTN.

    “In fact, assuming that a particular label sequence is given as the”

  • The paper relies on this without support.

    No reliable segmentation heuristic exists for cursive handwriting.

    “The SDNN architecture seems particularly attractive for recognizing cursive handwriting where no reliable segmentation heuristic exists.”

  • The paper relies on this without support.

    With real training data, the correct sequence of labels for a string is generally available, but the precise locations of each corresponding character in the input image are unknown.

    “With real training data, the correct sequence of labels for a string is generally available, but the precise locations of each corresponding character in the input image are unknown.”

  • The paper relies on this without support.

    There has been no proposal for a systematic approach to multilayer graph-based trainable systems.

    “However, there has been no proposal for a systematic approach to multilayer graph-based trainable systems.”

  • The authors state this as an assumption.

    The function used by fprop to compute the values attached to its output arcs is differentiable.

    “this method assumes that the function used by fprop to compute the values attached to its output arcs is differentiable.”

  • The paper relies on this without support.

    Natural handwriting is often a mixture of different styles, i.e., lower case printed, upper case, and cursive.

    “Natural handwriting is often a mixture of different “styles,” i.e., lower case printed, upper case, and cursive.”

  • The paper relies on this without support.

    Characters taken in isolation can be very ambiguous, but considerable information is available from the context of the whole word.

    “Characters taken in isolation can be very ambiguous, but considerable information is available from the context of the whole word.”

  • The paper relies on this without support.

    The intent of the writer is to produce a legible image.

    “Since the intent of the writer is to produce a legible image, it seems natural to preserve as much of the pictorial nature of the signal as possible, while at the same time exploit the sequential information in the trajectory.”

  • The paper relies on this without support.

    Verifying the amount on a check is extremely time and money consuming for banks.

    “The verification of the amount on a check is a task that is extremely time and money consuming for banks.”

  • The paper relies on this without support.

    There is a very high interest in automating check verification as much as possible.

    “As a consequence, there is a very high interest in automating the process as much as possible (see, for example, [112]- [114]).”

  • The paper relies on this without support.

    Even a partial automation of check verification would result in considerable cost reductions.

    “Even a partial automation would result in considerable cost reductions.”

  • The paper relies on this without support.

    The threshold of economic viability for automatic check readers, as set by the bank, is when 50% of checks are read with less than 1% error and the other 50% are rejected.

    “The threshold of economic viability for automatic check readers, as set by the bank, is when 50% of the checks are read with less than 1% error.The other 50% of the check being rejected and sent to human operators.”

  • The paper relies on this without support.

    Checks contain at least two versions of the amount: the courtesy amount written with numerals and the legal amount written with letters.

    “Checks contain at least two versions of the amount.The courtesy amount is written with numerals, while the legal amount is written with letters.”

  • The paper relies on this without support.

    On business checks, amounts are generally machine-printed, relatively easy to read but quite difficult to find due to the lack of standard for business check layout.

    “On business checks, which are generally machine-printed, these amounts are relatively easy to read but quite difficult to find due to the lack of standard for business check layout.”

  • The paper relies on this without support.

    On personal checks, amounts are easy to find but much harder to read.

    “On the other hand, these amounts on personal checks are easy to find but much harder to read.”

  • The paper relies on this without support.

    Feature extraction is traditionally a fixed transform generally derived from expert prior knowledge about the task.

    “Feature extraction is traditionally a fixed transform, and it is generally derived from some expert prior knowledge about the task.”

  • The paper relies on this without support.

    Traditional recognition systems rely on many handcrafted heuristics to isolate individually recognizable objects.

    “Traditional recognition systems rely on many handcrafted heuristics to isolate individually recognizable objects.”

  • The authors state this as an assumption.

    The standard deviation of the weighted sum scales linearly with the number of inputs if the inputs are highly correlated.

    “We chose to assume the second hypothesis since some units receive highly correlated signals.”
35 of 35

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Confidence discussion reverses the loss function's logarithmic sign

    Minor

    The paper defines the posterior probability as the exponential of minus the discriminative forward loss. Its later description of that loss as an estimate of the logarithm of the probability omits the negative sign.

  • Forward-penalty ordering is stated in opposite directions

    Minor

    The paper says a graph's forward penalty is larger than its subgraph's, but elsewhere says combining paths makes the overall penalty smaller than the penalties of individual paths. The latter also agrees with its statement that a constrained graph has a larger forward penalty than the unconstrained graph.

2 of 2

Citations

Cites

Flagged first, then by citations

Show all 129 references

Research works printed but not linked: 27

  1. matan, 1992, Neural Information Processing Systems, 4
  2. lecun, 1987, Mod�les Connexionnistes de l'Apprentissage (Connectionist Learning Models)
  3. bottou, 1991, Advances in neural information processing systems, 3
  4. keeler, 1991, Neural Information Processing Systems, 3, 557
  5. pereira, 1997, Finite State Devices for Natural Language Processing
  6. ott, 1976, Construction of quadratic polynomial classifiers, Proc IEEE Int Conf Pattern Recognition, 161
  7. cortes, 1994, Advances in Neural Information Processing Systems 6, 327
  8. lippmann, 1987, Neural-net classifiers useful for speech recognition, Proc IEEE 1st Int Conf Neural Networks, 417
  9. driancourt, 1991, MLP, LVQ and DP: Comparison & cooperation, Proc Int Joint Conf Neural Networks, 2, 815
  10. haffner, 1992, Advances in neural information processing systems, 4, 579
  11. simard, 1993, Advances in neural information processing systems, 5
  12. lecun, 1997, Reading checks with graph transformer networks, Proc IEEE Int Conf Acoustics Speech Signal Processing, 1, 151
  13. bengio, 1996, An EM algorithm for asynchronous input/output hidden Markov models, Proc Int Conf Neural Information Processing, 328
  14. rahim, 0, Discriminative feature and model design for automatic speech recognition, Proc Eurospeech '97
  15. wolf, 1994, Advances in neural information processing systems, 6, 745
  16. mohri, 0, Weighted determinization and minimization for large vocabulary recognition, Proc Eurospeech '97, 131
  17. gilloux, 1993, Recognition of cursive script amounts on postal checks, Proc Europ Conf Postal Technol, 705
  18. bourlard, 1989, Advances in neural information processing systems, 1, 186
  19. bengio, 1992, Advances in neural information processing systems, 4, 175
  20. minsky, 1961, Learning in random nets, Proc 4th London Symp Information Theory, 335
  21. rumelhart, 1986, Parallel Distributed Processing Explorations in the Microstructure of Cognition, i, 318
  22. denker, 1995, The Mathematics of Induction
  23. lecun, 1985, A learning scheme for asymmetric threshold networks, Proc of COGNITIVA'85, 599
  24. kramer, 1988, Advances in neural information processing systems, 1, 40
  25. parker, 1985, Learning-logic
  26. bottou, 1991, Une approche th�orique de l'Apprentissage Connexionniste Applications � la reconnaissance de la Parole
  27. schenkel, 1993, Advances in neural information processing systems, 5, 723

Cited by

Sorted by citations

See all 61,544

Related

See all

Metadata

Authors
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P.
Affiliations
  • AT&T (United States) (US)
  • Université de Montréal (CA)
ORCID
Venue
Proceedings of the IEEE
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
1998
Volume
86
Issue
11
Pages
2278-2324
ISSN
0018-9219 (ISSN-L)
DOI
10.1109/5.726791
OpenAlex
W2112796928
Semantic Scholar
162d958ff885f1462aeda91cd72582323fd6a1f4
Cited by
58,805
References
129
Citations by year
2026: 3932 · 2025: 6037 · 2024: 5859 · 2023: 5924 · 2022: 5751 · 2021: 6779 · 2020: 6359 · 2019: 5708 · 2018: 4169 · 2017: 2778 · 2016: 1818 · 2015: 1223 · 2014: 644 · 2013: 396 · 2012: 267
FWCI
12.51
Type
journal article
Language
en
Open access
Licence
Locations
  • Proceedings of the IEEE (publishedVersion)
  • HAL (submittedVersion, other-oa)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Handwritten Text Recognition Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Vehicle License Plate Recognition
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
Keywords
document recognition, handwritten character recognition, handwritten digit recognition, convolutional neural networks, graph transformer networks, global training, backpropagation, online handwriting recognition, field extraction, language modeling

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review