Deep Learning (DL) networks are the current great hope for artificial intelligence. They have achieved impressive feats like approaching or surpassing human performance on image recognition (including facial recognition) and language translation. They also fail in ways that raise equity and security concerns. The focus here is on so-called adversarial examples, which are prompts (images, texts, etc.) for which DL networks make wildly incorrect judgments if the prompt is perturbed in a very minimal way, such as changing a few pixels in an image. In evaluating how we ought to react to this challenge to DL’s ability to perform perceptual tasks, a comparison is made to human susceptibility to perceptual illusions, and philosophical accounts of robustness are consulted.

Email surveill@queensu.ca to receive the Zoom link.