The overarching goal of my work is to understand and create intelligent systems.

What are the fundamental mechanisms of human intelligence? How can intelligence be broken down and quantified? What are the limits of human intelligence and of intelligent systems in general?

I believe that many of the most amazing abilities we observe in humans and animals emerge naturally from simple yet incredibly powerful processes based on effective representations of the world. Thus, rather than explicitly working towards systems capable of highly sophisticated behaviors, my approach focuses on understanding the foundations of intelligence with the expectation that interesting insights and abilities will follow naturally. To pursue this goal, I build functional models in the form of computer algorithms. I then evaluate the models in the same way that humans evaluate themselves and each other: by placing them in games carefully designed to test their various capabilities.

How did intelligence arise in this universe? Was the emergence of intelligence inevitable? Under what conditions can and will intelligence come about? Under what conditions will social, cooperating agents arise? How does the computational model of the universe itself relate to the potential for intelligence?

As I study models of intelligence in virtual worlds, it occurs to me that my research is a bit... recursive. I wonder how we came to be here, studying ourselves, in a universe we know little about. Although I may not be able to peek outside, I can certainly look deeper inwards. I believe that simulated models of reality can shed light on many of our greatest mysteries, such as the emergence of complex intelligent life, the basis of ethics and morality, and the relationships between physics, computation, and intelligence.

Clearly, the above is only a high-level view. More specific information about each of my public projects can be found below. Although I have many other projects both ongoing and planned for the future, I prefer to keep the details of each private until I have concrete results to share.


Object-Oriented Re­inforce­ment Learning

Reinforcement learning offers a promising path towards the creation of intelligent agents, but state-of-the-art methods are woefully incapable when compared to even simple biological organisms. One of the reasons for this is the lack of object-based, compositional reasoning in deep-learning systems. In this project, we use the principle of object-oriented (i.e., relational) reasoning to design algorithms that learn and act more effectively.

External-Memory Streaming

Interaction with external-memory devices, such as hard disk drives (HDDs), is a significant challenge for modern applications due to the large delays incurred by random access. In this project, we develop and study a new model for external-memory applications called bowtie streaming. Our optimized I/O scheduling algorithms are able to maintain throughputs several times higher than the fastest prior methods.


QORA: Zero-Shot Transfer via Interpretable Object-Relational Model Learning

Gabriel Stella, Dmitri Loguinov. ICML, July 2024.

On High-Latency Bowtie Data Streaming

Gabriel Stella, Dmitri Loguinov. IEEE BigData, December 2022.