Enabling Autonomous Vehicles to Train Themselves
May 6, 2016
Enabling Autonomous Vehicles to Train Themselves
The article “How Deep Learning Networks Can Use Virtual Worlds To Solve Real World Problems” introduced me to the idea of augmenting a real world data set with data from a virtual world. This allows one to “what if” an infinite number of scenarios without having to attempt to find or create the situation in the real world data. Leveraging this, a system can be built to teach itself (in the spirit of AlphaGo) and could lead to great advances in automous vehicle safety in sooner-than-expected timeframes. This system could also be used to determine how sensor quality and changes in position or number of sensors would affect the resulting driving.
Using AI to Improve Synthetic Imagery
Reading “How Deep Learning Networks Can Use Virtual Worlds To Solve Real World Problems” and “Artificial Intelligence Can Now Design Realistic Video and Game Imagery” got me to thinking about how we can use AI to improve synthetic imagery. Rather than simply “create high-quality videos or images from low-resolution ones”, instead use AI to tunnel through the uncanny valley. Allow the algorithm to learn the difference between a synthetic (virtual) world and a real one and automatically fill in the gaps as needed. Providing this as a configurable step in the graphics pipeline may provide for never-seen-before photo-realistic worlds.
Newer
Older