AI & Indonesian Cultural Heritage
A batik motif is not random decoration. It has a grammar.
A supporting track. Smaller in volume than the Sci-ML work, but the question is a sibling of it: how a structure we already recognise gets carried into a model, and what falls out along the way.
Cloth motifs and script strokes both carry rules nobody ever wrote down. A model can imitate the shapes; capturing the rules is much harder, and that is where the work is.
An early attempt at style transfer on batik produced something that looked like batik from across the room and fell apart the moment you stepped closer. That failure was the interesting part. What the model missed was not colour or texture but rule: repetition, symmetry, how one ceplok unit meets its neighbour. Rules the artisan holds without ever writing them down.
From there the direction shifted to generative models: variational autoencoders for motif synthesis, with a sharper question underneath: what is actually stored in the latent space when a tradition is compressed into a few hundred numbers?
The second branch is script. With students, this track works on handwritten recognition for Sundanese, Javanese, and Balinese script, and machine translation for Kaili. This is low-resource work in the full sense: the data is scarce, the speakers are ageing, and there is almost no benchmark to borrow from the English-language literature. That is precisely why it is urgent.
Generative batik
Neural style transfer and variational autoencoders for batik motifs: learning repeating structure rather than mere texture.
Indonesian script recognition
Sundanese, Javanese, and Balinese script: CNNs, transfer learning, Faster R-CNN, and how image preprocessing shapes handwriting accuracy.
Regional-language NLP
Indonesian-Kaili machine translation and question generation in Indonesian: low-resource problems that rarely get attention.
The Roadmap, and Where You Fit
Every large direction has a version a student can take on: a whole, realistic project that also lays one brick in a long-running research programme.
Generative models for batik motifs
Can a model learn a motif's grammar rather than just its look? The style transfer and VAE work on this track shows exactly where the limit sits.
Push from VAE to stronger architectures, such as diffusion models or GANs with symmetry constraints, or design evaluation metrics that measure actual motif validity instead of pixel similarity.
Script and regional-language recognition
How far can Indonesian script recognition go on this little data? Sundanese, Javanese, and Balinese are covered, and much remains untouched.
Extend to scripts nobody has touched, build a new dataset (a contribution that outlives the thesis), or test whether modern low-resource methods genuinely help here.
Publications on This Track
A selection from the full list on Google Scholar.
Initial study of batik generation using variational autoencoder
A generative model trying to capture a batik motif's grammar rather than just its look, and the question of what is stored when a tradition is compressed into a few hundred numbers.
A study of batik style transfer using neural network
A style-transfer attempt made valuable by its failure, showing that what the model missed was the motif's rules, not its colour or texture.