Lecture, Conference Presentation
MIT 6.S094: Deep Learning for Human-Centered Semi-Autonomous Vehicles
- Deep learning systems are expected to progress toward using only 1% to 0.1% of human-labeled data, reserving human annotation for hard cases involving partial occlusions or extreme lighting while the machine annotates static or routine frames.
- Machine learning may eventually achieve 100% accuracy as the human annotation fraction increases from 0% to 10%, with future possibilities of reaching a state requiring no human annotation at all.
- The industry is transitioning from fully supervised to unsupervised methods, a process that will continue to require the creation of billions of frames annotated with ground truth for driver state algorithms.
- Neural network performance is anticipated to improve significantly in many important cases as depth increases, even without an increase in available data volume.
- Automakers are expected to adopt driver-facing cameras in every vehicle to generate the billions of miles of driver-facing data needed alongside existing forward-facing data for full automation.
- The market context, characterized as "pickup truck country," is likely to necessitate gradual steps toward full automation to accommodate a preference for manually controlled vehicles.
- Trust between machines and humans is predicted to develop gradually as the machine demonstrates awareness of the biological entity it is transporting.
- A specific 3D convolutional neural network processing 90 frames at 15 frames per second is expected to predict cognitive load into one of three classes: low, medium, or high.
- Code for the cognitive load prediction system is expected to be released online for processing webcam video streams in faster than real time.
- The industry still lacks a complete understanding of how complex structure and knowledge representation emerge from hundreds of millions of parameters in deep networks.
- It is hypothesized that neural networks may eventually be able to reason if the mechanisms behind their current emergent properties are understood.
- A self-driving car project receiving over 2,000 submissions is expected to continue for a period while researchers seek optimal solutions and prepare a journal paper.