Reconstruction Error Ratios for Prototype-Anchored Unsupervised Learning in Optical Character Recognition

Annotating rare alphabets is expensive: the people who can read them are few, and the collections that need them are large. PAUL (Prototype-Anchored Unsupervised Learning) starts from what such an expert can realistically provide — one to five labelled exemplars per class — and grows a labelled set from there.

Each class gets its own autoencoder, anchored on its exemplars. Candidates from an unlabelled pool are assigned by comparing reconstruction errors across class models: the ratio between the best and the runner-up reconstruction error decides whether a sample is confidently claimed or left alone. The pseudo-label set expands over iterations, and each expansion sharpens the class models.

We evaluate on EMNIST, KMNIST and two newly introduced Omniglot-derived benchmarks against SCAN, DEC, IDEC, JULE and centroid/GMM baselines, across DINO, CLIP and convolutional features. On EMNIST Balanced, PAUL reaches 66.9 ± 1.6% accuracy where the strongest baseline reaches 53.3 ± 0.9%.

The implementation is available on GitHub.

Code