Explainable and Robust AI for VILMA Virtual Laboratory

Explore explainable and robust AI applications in atmospheric science through this 39-minute conference talk by Kai Puolamäki, Associate Professor at the University of Helsinki. Gain insights into the Virtual Laboratory for Molecular Level Atmospheric Transformations (VILMA) Centre of Excellence and understand the importance of explainable AI and uncertainty quantification in atmospheric research. Discover SLISEMAP, a supervised manifold visualization method for explainable AI that provides local explanations for data items and creates two-dimensional global visualizations of black box models. Learn about challenges in molecular-level modeling, concept drift detection, and the exploration of AI models in scientific contexts. Delve into the motivations behind local explanations and dimensionality reduction as tools for scientific discovery. Examine practical applications of SLISEMAP, including predicting fuel consumption in cars, and understand its usage, problem definition, and potential impact on atmospheric science research.

Intro
Aim: understand which gases form completel new particles in the atmosphere
Challenge 1: molecular level modelling works, but wo require calculations longer than the age of the univers
Challenge 2: difficulties and biases in detectio relevant chemical species and molecular clust
VILMA VIRTUAL LABORATORY: AI FOR SCIENCE
CAN WE TRUST OUR PREDICTIONS? DETECTING CONCEPT DRIFT
HOW TO BUILD AND EXPLORE MODELS? XIPLOT
CRASH INTRO TO EXPLAINABLE AI (XAI): GLOBAL VS LOCAL EXPLANATIONS
MOTIVATION: LOCAL EXPLANATIONS
MOTIVATION: DIMENSIONALITY REDUCTI - AS A TOOL FOR SCIENCE
HOW DO THE MACHINE-LEARNING MODELS W SLISEMAP FOR EXPLAINABLE AI (XAI)
SLISEMAP: RANDOM FOREST PREDICTIN FUEL CONSUMPTION OF CARS
SLISEMAP: THE EFFECT OF RADIUS
SLISEMAP: MULTIPLE EXPLANATIONS
SLISEMAP: SUBSAMPLING
SLISEMAP: USAGE
SLISEMAP: SUMMARY
SLISEMAP: PROBLEM DEFINITION