SSL Autoencoder MNIST

Using self-supervised learning to pre-train an autoencoder on the MNIST dataset using 99% of available training data and fine-tuning the encoder using remaining 1% of labeled training data with a classifier output layer. Comparing the accuracy with a pure supervised classifier model trained on the same 1% of the training data.

Requirements

• Python 3.11

pip install -r requirements.txt

Notebooks

Notebook containing the code and results for the autoencoder and classifier is located in the notebooks folder.

Models

The trained models are located in the models folder. The autoencoder model is named autoencoder.keras and the fine-tuned model is named fine_tuned_model.keras. The supervised classifier model is named supervised_classifier.keras.