MNIST handwritten digit classification

1. Industrial challenge

Digit recognition is a compact introduction to visual pattern classification and document-processing systems. This example maps an already cropped digit image to one of ten classes; it does not locate digits in a page or read multi-digit fields.

2. Data set

The supplied collection contains 10,000 MNIST digit images at 28 × 28 pixels, with class counts matching the standard MNIST test collection. This project repartitions those images into training, validation and testing; it is not the standard 60,000-training / 10,000-test MNIST benchmark protocol.

Dataset source and access conditions. Source reviewed on 24 September 2026.

The saved augmentation settings have zero rotation and translation, with no random reflections. Images are read in class-folder order and sorted filename order within each class.

3. Model

The following configuration comes from the saved trained model. The same topology is used before and after training; no separate architecture-selection run is recorded.

Sigmoid or softmax outputs are uncalibrated model scores. The predicted class is the label with the largest output score.

4. Training strategy

Training minimizes cross-entropy using Adam. The saved configuration uses a learning rate of 0.001, mini-batches of 32 and no regularization.

The summary table and epoch history are retained as separate native report outputs. Epoch-history endpoints need not equal the restored best-validation model; they are not additional test metrics.

5. Model selection

No architecture selection or feature selection experiment is recorded. The same network topology is used before and after training. The validation subset monitors training; the reported evaluation uses the saved testing assignment.

A constant classifier predicting the most frequent training label (one) achieves 10.65% accuracy on the same testing images, compared with 96.10% for this model. The constant label is chosen from the training images only.

6. Testing analysis

Testing uses the 2,000 images assigned to the testing subset in the saved project. The confusion matrix and class metrics agree with exported inference, and repeated reloads preserve the same evaluation images. Rows below are actual labels; columns are predicted labels.

The model predicts a single digit from a cropped image. The largest class-level recall shortfall is for nine (0.919). These results use this project’s own partition of the 10,000-image collection, not the standard MNIST training/test protocol.

7. Model deployment

Recognizing handwritten digits

See how the model recognizes two handwritten digits. Each example pairs the original image with the predicted digit and the scores for all ten classes.

Example 1

Predicted class: one
Model score: 99.97%

Class	Output score
eight	0.000007
five	0.000000
four	0.000011
nine	0.000001
one	0.999687
seven	0.000289
six	0.000000
three	0.000002
two	0.000004
zero	0.000000

Example 2

Predicted class: four
Model score: 93.20%

Class	Output score
eight	0.064999
five	0.000000
four	0.932005
nine	0.002492
one	0.000000
seven	0.000375
six	0.000088
three	0.000000
two	0.000037
zero	0.000003

These are illustrative predictions, separate from the testing metrics above. Output scores have not been calibrated.

Download the original Neural Designer project and native figures, then relink the source dataset. The ZIP is a project archive; it does not contain the standalone Neural Engine runtime.

For batch inference, export a deployment package from Neural Designer. Keep the generated Python wrapper beside its engine/ and model/ folders. The supplied export uses Windows x64 and Python 3, with CPU inference by default.

python mnist.py sample.bmp

Preserve the trained vocabulary and sequence limit for text, or the saved image dimensions, channel handling and scaling for images. Record the model version and input quality before sending outputs for human review. The wrapper’s score output is not evidence of probability calibration.

8. Scope and limitations

Handwriting, framing, background polarity and image quality can shift outside this collection. Digit localization, rejection of non-digits and full-page optical character recognition require additional components. Writer-level separation is not documented.

The supplied artifacts do not establish independence by person, product, author, acquisition session or site. Check duplicates and group-related records before reporting performance on a new population. Calibration and external validation are not established; monitor drift and review errors before operational decisions.

References

• MNIST handwritten digit classification — dataset source
• Supplied Neural Designer project and native figures, 24 September 2026 (ZIP)
• More Neural Designer examples