Veos Digital’s Artificial Intelligence and Applied Mathematics Laboratory investigates mathematically grounded, intelligible artificial intelligence solutions to improve and promote the interplay between artificial and biological agents in decision-making.
We interact with perceived reality through complex, nonlinear transformations computed at impressive speed by our brain and biased by our experiences, memories, and genetics. The interaction of these factors gives rise to intelligence, learning, and creativity.
Through observation and probing of biological agents, neuroscientists aim to measure and quantitatively explain perception, intelligence, learning, and creativity. We turn our attention toward artificial intelligence to investigate the same concepts employing completely observable artificial agents.
We aim to describe learning machines in mathematically sound environments. A robust mathematical formalization allows us to implement artificial intelligence methods from a principled perspective, letting theorems guide algorithmic choices and drive our applied efforts.
Finally, we investigate hybrid decision-making, where biological and artificial agents are mutually involved in decision processes. We believe that this approach allows us to reach original, valid solutions in well-known and novel tasks and shed light on the mechanistic understanding of intelligence, learning, and creativity at the biological level.
Projects
Parametric Machines
One of the most obscure processes involved in designing deep learning applications is the choice of the network’s architecture. We define mathematical spaces where neural networks live and can be combined and interpolated in original ways, following formal definitions and theorems guaranteeing their optimality.
Currently applied to:
- Automatic machine learning
- Time series classification and forecast
- Vision
- Parsimonious encoding in Variational methods (VAE, RVAE, CVAE)
Explainability
Understanding what and how a machine learns from a dataset is crucial for developing robust machine-learning-based applications. We design and borrow techniques from Functional Analysis and Topology to maximize the explainability of deep learning methods. Functional Analysis allows us to define principled regularization strategies, while Computational Topology provides human-readable descriptions of learned transformations and decision points.
Currently applied to:
- Hybrid-decision making in time series prediction
- Assisted medical diagnosis
- Transparent recommendation systems
- Feature selection in highly sparse datasets