Deep Learning Exercises

The problem

Two practical exercises: predict COVID-19 case numbers from survey features (regression), and classify food photographs into eleven categories (Food-11) with limited labelled data.

What was built

• 01Regression: a fully connected network trained with a custom TensorFlow training loop, MSE loss with L2 regularisation, and model selection on a validation split.
• 02Classification: a ResNet-18 architecture written from scratch in Keras (no pre-trained weights), with data augmentation.
• 03A pseudo-labelling step to make use of the unlabelled images in the dataset.

My role

Individual coursework, adapted by the course instructor from the ML2021 assignments by Hung-yi Lee's group.

Technologies