Dropout and Weight Regularization Experiments Notebook

Overfitting is the default fate of neural networks trained on limited data, and regularization is the cure — but learners rarely get to watch the cure actually work. This project builds a notebook laboratory where dropout and weight penalties are applied, varied, and observed rather than memorized.

Working in Jupyter with Keras and TensorFlow, the intern constructs a small neural network prone to overfitting by design, then runs controlled experiments: adding dropout layers at different rates, applying L1 and L2 weight regularization at different strengths, and combining the techniques. For every configuration the notebook plots training and validation loss together, so the divergence signature of overfitting — and each technique's effect in narrowing that gap — is directly visible. The experiments are organized so users can modify parameters and rerun cells to test their own hypotheses about what will happen next.

The intern comes away understanding not just what regularization is but how each technique behaves and how much of it is enough, along with experience structuring reproducible notebook experiments — the same discipline that underpins credible deep learning research.