Pouya Shati Homepage

Research Summary

The recent advancements in machine learning have revolutionized almost every aspect of our lives. Given the magnitude of this transformation, it is crucial to remain vigilant about the potential ways in which AI can act contrary to human interests. However, many modern solutions rely completely on large and complicated black-box models that are challenging to understand or formally interact with.

Two important factors in producing human-compatible and trustworthy machine learning solutions are interpretability and constrainedness. Interpretable models are compact and easy to reason about by humans and algorithms. Constrained machine learning aims to specify and enforce formal specifications on solutions during training or inference. Both enable safety-critical applications, analysis and guarantees with regard to privacy and security concerns, and integration of audits by domain experts.

My research is focused on using exact methods and other formal components from classical AI for interpretable and constrained machine learning. Exact methods are an umbrella term referring to a family of paradigms aimed at solving computationally challenging problems. I use these symbolic components independently or in conjunction with conventional ML to develop human-compatible solutions in a variety of well-known problems. I show that the complex reasoning capabilities and theoretical guarantees of exact algorithms can complement the impressive scaling of deep learning without significant cost to solution quality.