Discovering the World Through Unsupervised Statistical Relational Learning

Machine learning is popular nowadays, thanks to the impressive results achieved by systems like DeepMind’s AlphaGo, OpenAI’s language prediction model GPT-3 or Amazon’s speech recognition system Alexa. At the basis of these successes, there is representation learning, which enables training deep neural networks in an unsupervised fashion and provides the starting conditions for subsequent task-specific training. However, current representation learning strategies use large neural networks and consume large amount of data, thus being data and energy inefficient. In contrast, humans learn from limited data in a very efficient way. This is due to the fact that humans are able to perform reasoning, while representation learning strategies lack such capability. This research project aims to overcome these limitations by providing the mathematical foundations for the integration between unsupervised learning and reasoning AI systems based on logic. Specifically, the aim is to devise algorithms enabling the discovery of symbolic representations from noisy/ambiguous data together with their relations and being able to adapt the acquired relational knowledge over time. The resulting solutions will be applied to improve image understanding in autonomous driving and to gain insights about causal reasoning and learning symbolic abstractions in mathematical domains.