Data-driven challenges in AI: scale, information selection, and safety
summary
Proposes a new decision-tree objective that yields logarithmic prediction time and is optimized with SGD, demonstrating superior precision and prediction speed on a multi-label task with 200,000 classes; replaces expensive fully‑connected outputs in CNNs with the tree to jointly learn features and labels. Introduces a reconfigurable gating scheme for a 1:6 scale autonomous car with three cameras plus LiDAR that activates only relevant feature extractors, cutting FLOPs to near single‑sensor cost while retaining multi‑sensor performance. Implements a GAN-based online monitor whose discriminator optimum vs. actual steering flags sensor–actuator mismatches in real video demos.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:25Questions
- 2:25Multiclass classification
- 5:15Objective function
- 11:52Autonomous car driving
- 14:00Gating mechanism
- 17:50Comparison
- 19:30Safety
- 24:12Summary
related talks
28mChallenges & future trends of Deep Learning
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
24mKeynote: Next-generation frameworks for Large-scale AI
29mDeep Learning, Neuroscience and the future of AI
31mDeep learning for self-driving vehicle technology
40mYou can AI like an Expert
41mSpotlight on AI use cases and innovators around the world
25mGenerative AI vs Adaptive AI - Applications and Algorithms