AI with AI
Episode 2.37: The Call of Aftershocktopus (Part 2)
Researchers at Berkeley, Washington, and Chicago identify “natural adversarial” examples that cause classifier accuracy to significantly degrade, likely due to an over-reliance on color, texture, and background cues. Andy and Dave then discuss a series of events following a Nature paper on application of deep learning to aftershock patterns of earthquakes, wherein other researchers raised questions on the researcher (one demonstrating that a simple logistic regression does better; and another showing that the original researchers included their test data set in their training data set). A new study by the Insurance Institute for Highway Safety shows that drivers overestimate the capability of vehicle automated systems, with Telsa’s Autopilot leading the rest in overestimation. Goodfellow, Bengio, and Courville publish their 800 page tome on Deep Learning. The Classic Paper of the Week comes from Pattie Maes and Rodney Brooks, who published Learning to Coordinate Behaviors in 1990. The video presentation of the octopus research makes the video of the week. NASA streams 24/7 with OUTERHELIOS, a neural network trained on Coltrane to produce non-stop free jazz (though the feed may now be “static only”).