Neural Networks Projects
Portfolio Experiments

Hands-On NN Projects

Reading tutorials is not the same as training models end-to-end. These project ladders go from quick wins to portfolio pieces. For each project, aim for: a clean train/val split, logged metrics, a short README (data, commands, results), and fixed random seeds when comparing ideas.

baseline first version data ablations

Beginner

  • MNIST / Fashion-MNIST — MLP then a small CNN; compare accuracy and training time.
  • CIFAR-10 — data augmentation, residual-ish block or pretrained backbone.
  • Tabular classification — embeddings + MLP vs gradient boosting (XGBoost/LightGBM) as a sanity baseline.

Intermediate

  • Transfer learning — classify a small custom image set with a frozen then fine-tuned ResNet or ViT.
  • Text sentiment — fine-tune a small transformer or use TF-IDF + linear model as baseline.
  • Time series — LSTM or 1D CNN for forecasting; report MAE/RMSE on a rolling validation.
Ship a minimal demo: Gradio, Streamlit, or a notebook with clear “Run all” instructions—recruiters can reproduce your work.

Summary

  • Start with baselines; improve with architecture, data, and regularization—not random tweaks.
  • Document setup and results; treat each repo as a mini research log.
  • Next: practice exercises and quizzes to test understanding.

Reinforce concepts with practice exercises in the next page.