Summary 🎨

This final project for Neural Networks 2023 reimagined a DCGAN as a playground for generating Lissajous-inspired visual motifs.

The experiment 🧪

• I started with a classical DCGAN and tuned the generator to output sequences of points that resemble Lissajous curves.
• The synthetic dataset was built on top of the Harmonumpyplot script by @tuxar-uk, then refined with custom preprocessing so the visuals stayed sharp.
• Loss balancing was a dance between stability and visual richness: the discriminator needed subtle annealing to avoid mode collapse.

Results & notes 📝

• The generated assets are available in the linked GitHub repo, along with the dataset and Colab notebook that helped me iterate quickly.
• It became a testbed for exploring how structured, periodic signals interact with adversarial training.

What I learned ✨

• GANs are still sensitive to initialization, but curating the dataset to match the induction biases of a Lissajous curve pays off.
• The project also taught me the value of quick visual feedback loops: seeing each epoch output helped diagnose divergence earlier than loss curves alone.

👾 See you space cowboy.