Our First data challenge

This is a test. The LSU Interdisciplinary AI-JC attempting the first hands-on activity.

Distinguishing Ancient Chinese Written Characters

• Paper: https://www.nature.com/articles/s41597-024-02933-w
• GitHub (Oracle-MNIST): https://github.com/wm-bupt/oracle-mnist

Steps:

• Data pre-processing

  • Download the data: from this repo - Credits: Oracle-NMIST Team
  • Look at the data (inspect)
  • Process the data

• Data Set preparation

• NN set up

  • brainstorm on possible architectures
  • Start with the MNIST CNN (LeCun)
  • Modify to get better performances

---

Results

Qur final architecture and results are reported in the images below.

Our final test accuracy (defined as the fraction of correct predictions on the total sample) is : 90.87% (Pure LeNet was giving 80% after 60 Epochs)

Observations

• Switching from Sigmoid() to ReLU() initially caused the loss to explode to ~10⁴.

  • Root cause: images were still in uint8 format.
  • Fix: convert to float32 and normalize pixel values to the range [0,1].

• Pooling comparison:

  • Replacing AvgPool2d with MaxPool2d degraded performance. The model ran better with AvgPool2d, so we reverted the change.

• Dropout experiments to mitigate overfitting:

  • Dropout(0.2): produced a small improvement.
  • Dropout(0.5): significantly improved generalization — reaching ~90% validation accuracy.

What We Did Not Try

• Data augmentation (e.g., rotations, translations, flips, noise injection)
• Deeper architecture (e.g., additional convolutional blocks or more filters).

Both directions could further improve performance.