A masterclass in Mexico City
Dr. Omar Costilla Reyes announced a masterclass in Spanish, “Inteligencia artificial y técnicas de programación aplicadas en ciencia” (artificial intelligence and programming techniques applied to science), in March 2023 at Tecnológico de Monterrey’s Santa Fe campus in Mexico City. The announcement flyer carries the logos of NDS Cognitive Labs, Tecnológico de Monterrey, and the Global Information Technology Management Association (GITMA). Dr. Costilla Reyes was then a research scientist at MIT CSAIL (MIT Computer Science and Artificial Intelligence Laboratory). Afterward he thanked Tecnológico de Monterrey and Gustavo Parés for the invitation, wrote that he enjoyed interacting with the students and that they had an excellent discussion, and said he was glad to see excitement and curiosity among them and hoped it would drive positive uses of AI.
Interpretability and program synthesis
He described the talk as covering machine-learning interpretability and program-synthesis techniques that his team at MIT was working on, and framed it with a concern he stated plainly: large language models suffer from an inherent black-box nature, making them incapable of scientific insight into our world. A large language model is a neural network trained on very large amounts of text; its answers come from a very large number of numerical weights that a reader cannot follow step by step. Interpretability methods try to expose what a model has learned, and program synthesis, the automatic generation of programs from examples or specifications, produces solutions written as code that a scientist can read, check, and revise. The slide shown in the event photograph is titled “AI for code and science,” with the label IAP 2023 (MIT’s Independent Activities Period) and the MIT CSAIL logo, the same title as the MIT workshop he co-organized a month earlier.
Research context
The topic belongs to the Neurosymbolic Programming for Science project and its 2022 position paper “Neurosymbolic programming for science,” which describe how neural and symbolic components can be combined for scientific discovery. The masterclass also fits a pattern in his outreach to students in Mexico, which includes the Machine Learning for Mexico symposium in 2018 and the Tecnológico Nacional de México academic meeting in 2020; each has its own page on this site.