Correlations-Oriented Representation Learning

Deep learning methods have propelled many of the recent remarkable achievements of artificial intelligence, yet their inner workings remain enigmatic. How do these methods represent the training examples in their multiple processing layers? How are these representations used to make decisions on previously unseen data?

The Correlation-Oriented Representation Learning (CORaL) project delves into the intricate relationship between data correlations and the hidden representations of deep neural networks.

CORaL's core objectives encompass: i) characterizing the structure of data correlations in tasks where deep learning methods excel; ii) providing a theoretical description of the interplay between the learning dynamics of deep networks and data correlations; and iii) validating these theoretical insights on benchmark machine-learning datasets.

The Researcher will approach these objectives with an effective mix of numerical experiments, scaling descriptions and rigorous arguments, inspired by statistical physics and fostered by the unique interdisciplinary scientific community of SISSA.

By describing the impact of data correlations on representation learning, CORaL holds the potential to enhance the interpretability and efficiency of deep learning methods across diverse domains, while also enhancing our general understanding of the principles of learning from examples.