Introduction
This presentation, delivered by Jose M. Alvarez at the NVIDIA GPU Technology Conference in San Jose, 2019, explores the critical challenges and innovative solutions involved in scaling deep learning technologies for autonomous vehicles (AVs).
The focus is on developing robust, efficient, and reliable AI systems capable of handling the immense data requirements and complex environmental conditions inherent in self-driving technology. Key areas discussed include the development of large-scale datasets, active learning strategies, domain adaptation techniques to ensure performance across diverse scenarios, and methods for improving the accuracy and efficiency of deep neural networks (DNNs).
Key Topics Covered
- The Challenge of Scale: Addressing the vast data collection, labeling, and training needs for AV perception systems.
- Active Learning: Techniques to efficiently select the most informative data for training, reducing the overall data requirement while maintaining performance.
- Domain Adaptation: Methods to ensure AI models perform reliably across different environments, weather conditions, and lighting.
- Accuracy vs. Efficiency: Strategies for optimizing DNNs for both high performance and real-time deployment, including over-parameterization, joint training and pruning, and filter decomposition.
- NVIDIA AI-Infra: The role of NVIDIA's infrastructure in supporting these advanced AI development workflows.
Significance
The insights provided highlight NVIDIA's advancements in enabling the development of sophisticated deep learning models essential for the future of autonomous driving. The presentation details research and methodologies aimed at making AV perception systems more accurate, robust, and scalable.
