Projects Proposals
Available Project Proposals
There are several possible projects on disparate topics, categorised below. Roughly, each first-level dot corresponds to a project.
If you are interested in any of the following project feel free to contact our researchers:
Explainable AI
Symbolic Knowledge Extraction (SKE)
- Design, prototype, and assess a novel symbolic knowledge extraction algorithm that supports recursive rules.
Symbolic Knowledge Injection (SKI)
- Study, design, and implement the [Lyrics] symbolic knowledge injection algorithm. An implementation of Lyrics already exists but it is quite old and not mantained.
- Study, design, and implement the [Fibred NN] symbolic knowledge injection algorithm.
- Study, design, and implement the [GRAM] symbolic knowledge injection algorithm.
- Study, design, and implement the [Knowledge-aware object detection] symbolic knowledge injection algorithm.
- Study, design, and implement the [Semantic loss function] symbolic knowledge injection algorithm.
- Design and implement the extension of the [PSyKI] library to support graph structured data and graph processing through deep learning frameworks (e.g., [Tensorflow-GNN]). Graph structured data must be supported both from the data domain (under the form of collections of graphs) and the knowledge domain (under the form of ontologies).
- Design, prototype, and assess a novel knowledge base quality metric that defines the coverage of a given knowledge base w.r.t. a given dataset. The proposed metric should be integrated in the [PSyKI] library.
- Design and implement the extension of the [PSyKI] library to support complex-structured datasets such as images, texts, etc. The extension should consider checking if the imported dataset type is compatible with the selected injection mechanism and its corresponding knowledge.
References
[Lyrics] Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Marco Gori. "LYRICS: A General Interface Layer to Integrate Logic Inference and Deep Learning"
[FIbred NN] Artur S. d'Avila Garcez, Dov M. Gabbay. "Fibring Neural Networks"
[GRAM] Choi, Edward, et al. "GRAM: graph-based attention model for healthcare representation learning."
[Knowledge-aware object detection] Fang, Yuan, et al. "Object detection meets knowledge graphs."
[Semantic loss function] Xu, Jingyi, et al. "A semantic loss function for deep learning with symbolic knowledge."
[PSyKI] https://github.com/psykei/psyki-python
[Tensorflow-GNN] https://github.com/tensorflow/gnn
Transfer Learning
Symbolic Transfer Learning
- Use the symbolic knowledge obtained by a SKE algorithm for a certain task in a another task (using a SKI algorithm). It is necessary to study how both SKE and SKI work, design and implement a running example.