Uniting Statistical Testing and Machine Learning for Safe Predictions

Descripción del proyecto

Solucionar el problema del efecto caja negra en el aprendizaje automático

El reciente crecimiento y avance de las tecnologías de aprendizaje automático han aportado importantes beneficios e interacciones con diversos sectores, repercutiendo en los sistemas autónomos, la toma de decisiones, los descubrimientos científicos y el diagnóstico médico. Sin embargo, a medida que los sistemas predictivos de aprendizaje automático se vuelven más sofisticados, su razonamiento suele ser más difícil de interpretar, lo que plantea problemas de seguridad. En este contexto, el proyecto SafetyBounds, financiado por el Consejo Europeo de Investigación, abordará este reto de seguridad, comúnmente denominado «efecto caja negra». En concreto, establecerá límites de error interpretables y precisos en las predicciones de aprendizaje automático, de los que se pueden derivar valiosos conocimientos. En última instancia, el proyecto sentará las bases para futuras prácticas de diseño de aprendizaje automático y sistemas de aprendizaje.

Objetivo

Recent breakthroughs in machine learning (ML) have brought about a transformative impact on decision-making, autonomous systems, medical diagnosis, and creation of new scientific knowledge. However, this progress has a major drawback: modern predictive systems are extremely complex and hard to interpret, a problem known as the black-box effect. The opaque nature of modern ML models, trained on increasingly diverse, incomplete, and noisy data, and later deployed in varying environments, hinders our ability to comprehend what drives inaccurate predictions, biased outcomes, and test time failures. Perhaps the most pressing question of our times is this: can we trust the predictions for future unseen instances obtained by black-box systems? The lack of practical guarantees on the limits of predictive performance poses a significant obstacle to deploying ML in applications that affect people's lives, opportunities, and science.

My overarching goal is to put precise, interpretable, and robust error bounds on ML predictions, communicating rigorously what can be honestly inferred from data. I call for the development of protective ecosystems that can be seamlessly plugged into any ML model to monitor and guarantee its safety.

This proposal introduces a unique interplay between statistics--the grammar of science--and ML--the art of learning from experience. Leveraging my expertise in both domains, I will show how statistical methodologies such as conformal prediction and test-martingales can empower ML, and how recent breakthroughs in ML such as semi-supervised learning and domain adaptation technologies can empower statistics. I will tackle challenges rooted in real-world problems concerning (1) availability and (2) quality of training data, as well as (3) test-time drifting data.

A successful outcome would not only lead to a timely and rigorous way toward safe ML, but may also significantly reform the way we develop, deploy, and interact with learning systems.