CARL presenting at the 39th Annual Education and Research in Computer Aided Architectural Design in Europe (eCAADe) Conference

Very excited to be presenting a paper titled: "Visualizing Deep Learning Models for Urban Health Analysis." Deep learning models are often referred to as “black-box” models because their inner-workings remain obscured behind hundreds of thousands, and sometimes millions, of parameters. The development of analytic methods to address this problem is currently a pressing problem (Yosinski et al., 2015; Zeiler and Fergus, 2014). This research proposes a new mixed-methods approach using Gram matrices (Gatys, Ecker, and Bethge 2016) to identify visual features in DL models that are correlated with health outcomes. Specifically, DL models are trained on satellite image datasets of US census tracts representing high and low rates of obesity, diabetes, and heart disease. These models are then analyzed, and the results assessed, providing urban planners, designers, and architects with a workflow to extract insight from DL models.