Reproduce “Neural Joking Machine : Humorous image captioning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kota Yoshida, Munetaka Minoguchi, Kenichiro Wani, Akio Nakamura, Hirokatsu Kataoka
An AI system called the Neural Joking Machine (NJM) was developed to generate humorous image captions, moving beyond traditional descriptive captioning. The NJM, trained on a large dataset of humor-rated image-caption pairs, generated captions that received an average of 3.23 stars on the Japanese Bokete website, nearly double the 1.71 stars of baseline systems.
Key insights
- Subjective human concepts like humor can be effectively integrated into machine learning models through quantified human feedback in training data.
- Modifying the loss function to directly optimize for human-perceived humor (the "Funny Score") can steer generative models towards creative, non-descriptive outputs.
- Large-scale datasets of user-generated, subjectively rated content are crucial for training AI systems on complex, qualitative human attributes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.