Drago Anguelov — Robustness, Safety, and Scalability at Waymo

Origins

Dragomir Anguelov recounts the origins of the inception architecture, highlighting his early work at Google on Street View and computer vision tasks. He explains how the initial deep learning models were met with skepticism and how the team transitioned from traditional methods to deep learning. The development of the inception module, with its innovative architectural tweaks, marked a significant milestone in computer vision 1 2.

   

Adversarial Examples

The discovery of adversarial examples was a pivotal moment in deep learning. Drago describes how the team encountered unexpected results when testing their models, leading to the realization that small changes in input could drastically alter outputs. This finding underscored the need for robust models capable of handling such anomalies 3 4.

   

Impact

The inception architecture's influence extends beyond its initial success, impacting subsequent deep learning models. Drago discusses the evolution of these models, noting the shift towards large transformer models and the continued push for higher accuracy. He reflects on how these advancements have significantly improved performance on benchmarks like ImageNet 5.