LISTIT (Learning non-Isomorph Structured Transductions for Image and Text fragments) is a 4-years research project funded by the Italian Ministry of Education and Research, under the SIR framework (contract n. RBSI14STDE). This is the official repository of the project referencing software libraries and code developed within the project and the associated pubblications.
The high level goal of LISTIT is the design and development of machine learning and deep learning methodologies generalizing supervised learning to structured samples both in input and output to the learning model. LISTIT considers primarily tree data but also targets more general classes of structures, including graphs with cycles.
LISTIT applications target natural language processing, image captioning, biomedical and life-sciences data.
LISTIT has built a range of learning models targeting
- Structure embedding : encoding of topology-varying input samples (trees, graphs) into fixed-size adaptive vectorial embeddings
- Structure decoding : generation of topology-varying structured predictions (trees, graphs), possibly conditioned on vectorial encodings of input structure