Data Lakes are emerging as data management infrastructures for storing data in various schemata and structural forms. Their goal is to serve as a single entry point for the data analysis process across highly heterogeneous datasets, supporting analytical tasks following a schema-on-read approach, in which data is discovered and integrated when it is to be used. Due to their semantic and structural heterogeneity, Data Lakes bring integration challenges to a new scale of complexity.
In this project we will explore the interface between emerging Deep Learning representation paradigms and heterogeneous dataspaces (structured, semi-structured and unstructured data), investigating how contemporary deep learning architectures and their induced embeddings can serve as a foundation for data integration, fusion and interpretation on data lakes. You will have the opportunity to design novel AI architectures exploring the space of contemporary methods such as transformers, variational autoencoders and graph neural networks.
Topics of interest include:
- Design of novel neural and variational embeddings for tables.
- Applications of table embeddings in inference tasks.
- Embeddings as a supporting paradigm for data fusion.
- (Semantically deep) program synthesis for data transformation (few-shot learning settings).
- Explaining table differences (via explainable neural architectures).
Applicants are expected to have:
- An excellent undergraduate degree in Computer Science or Mathematics (or related discipline), and preferably, a relevant M.Sc. degree.
- Confidence and independence in programming complex systems in Java or Python.
- Previous academic or industry experience in Natural Language Processing, Machine Learning or Data Science (desired).
- Excellent report writing and presentation skills.
Qualified applicants are strongly encouraged to informally contact Norman Paton ([Email Address Removed]) and Andre Freitas ([Email Address Removed]) to discuss the application prior to applying.