Building a Character-Level Movie Title Generator
After diving deep into automatic differentiation in my previous post, I was inspired to take on a new challenge: building a neural network that generates movie titles character by character. This project combines machine learning fundamentals with a fun objective of generating new movie names, and dives into understanding the intuitive reasoning of the results.
Sample Generated Titles
"David: A Quest of Saladin"
Historical/Adventure style
"Lost in Manhattan"
Classic rom-com style
"Speed Alien"
Action/Sci-fi mashup
"The Chocolate Cave Kid"
Family/Adventure style with a twist
These are actual outputs from our model, showcasing its ability to generate diverse and genre-aware titles!
Training Results
Generate Your Own Movie Title!
Want to see what the model can create? Click the button below to generate a random movie title. Some are quite creative, others are amusingly absurd, but they all show how the model learned patterns from real movie titles.
Takeaways
The most interesting part ended up being BatchNorm. I'd read about it before, but watching training diverge without it and then stabilize with it made the intuition click in a way that reading papers hadn't. Adding it to the hidden layers made the difference between the model learning anything at all and converging cleanly.
The outputs are a mix of surprisingly good and completely absurd — which is kind of the expected behavior for a character-level model on 45K titles. It learned title-like patterns without understanding what a movie is.
Code is on GitHub if you want to poke around or train your own.