From Uncertainty in Deep Learning to Data Efficient AI

Data-driven generative models and multi-agent simulators are often used to evaluate (back-test) trading strategies. However, learnable strategies can exploit imperfections in the data generators/simulators, converging to degenerate or trivial solutions. These are then wrongly evaluated as performing well under the flawed generative models and simulators.  

Instead of trying to label as much data as possible to cover all possible scenarios when training these data-driven generative models and simulators, this project aims to utilise uncertainty-aware models which know when the data-generators/simulators are driven out of distribution. It then uses an active learning paradigm to label new data only for these regimes using a few annotated samples.  

The project ties to two of JP Morgan Chase & Co.’s focused areas of research: Synthetic Data Generation and Simulated Multi-Agent Systems. In both, with advancements in fundamental methodology, this project will tackle new problems such as developing data-driven generative models and multi-agent simulators which cannot be exploited by learnable strategies, as well as developing learnable strategies which do not fall prey to imperfections in the data-generators and simulators.  

The project’s data efficient AI will reduce the amount of labelling required for such AI systems, enabling new and exciting applications which would have been impossible otherwise. Development of these new fundamental methodologies lies at the core of this project.