TECH

MIT and Microsoft are trying to solve "blind spots" in AI to help stand-alone cars

Methods for teaching a self-contained system are not always able to reproduce real-world situations, which leads to less effective learning. MIT and Microsoft are therefore developing a method to make autonomous driving systems "smarter".
The researchers explain that, in most cases, simulations are attempted that attempt to recreate all sorts of possible scenarios, including the least favorable scenarios. However, the virtual plan does not "reproduce" what is called "blind spots" - which can be translated as "dead spots" or "empty spots", in this case representing unexpected situations that may arise along the way - and the lack of knowledge leads the vehicle to react incorrectly to the problem.
The proposal of both institutions is that humans intervene directly in the learning process. The use of the simulations will continue, but there will be a person observing every "decision" of the system to prepare an analysis report on the errors made by the machine - or those that were about to happen.
The resulting information will be combined with "real-world" human-generated data, allowing the machine learning system to create a model to identify when more information about how to proceed is needed.
The method already validated in video games, but now the intention is to apply it in the "training" models of autonomous cars and robots. Sapo