Reimagining AI for a world on fire

Abstract

With the rise of machine learning (ML) and the vast availability of data in the web, artificial intelligence (AI) has quickly taken over the world, effectively changing the way we interact with the world and other humans. Not surprisingly, this shift in technology has spurred well-deserved discussions on AI's negative impact on employment, social justice, and personal privacy. Nonetheless, even though ML usually feeds on large-scale computing systems, few works have focused on reducing the environmental cost of training ML models --- a problem that is bound to scale . My objective in this project is to reduce the carbon footprint of ML while improving (or at least not degrading) the performance of the state-of-the-art. In turn, I expect to i) make the analysis of large databases more ecologically sustainable, and to ii) alleviate the ecological impact of using ML models in large-scale applications.This can be done by, e.g., devising learning algorithms that converge faster, creating lightweight predictive models, and selecting only a subset of data to train on. This project follows my recent successes in computationally-efficient training of Bayesian ML models from distributed data (Mesquita et al., UAI 2019; El Mekkaoui et al., UAI 2021; De Souza et al., AISTATS 2022), alleviating the cost of training Gaussian process latent-variable models (De Souza et al., AISTATS 2021), showing that some deep models for graph data can be severely simplified without any impact on performance (Mesquita et al., NeurIPS 2020), and analyzing the expressive power of teporal graph networks (Souza et al., NeurIPS 2022). (AU)