1

In an AI application (for example, self-driving), there are usually many different reading devices/sensors to ensure the outcome is correct. More specifically, a self-driving car can use object tracking with cameras, road-integrated optic fiber, sound analysis, and so on.

In many cases, these readings can be very difficult to integrate into a single model and get a single output. Assume each device provides an independent output. What are some ways that I can do to combine them and find out the most likely readings?

For instance, device A says there are 10 people standing 5 meters away, B says there are 5 people standing 10 meters away, and C says there are no people at all. The easiest way is to do a weighted average/voting, but it can make more correct readings less effective. If A is totally precise 95 percent of the time, no matter how much weight I assign to A, it will be affected by the less accurate ones.

seermer
  • 111
  • 1
  • First of all, I would exclude correlated inputs. If input data has a completely different format, you can feed it using different heads. I may also want to normalize it if possible – Aray Karjauv Jun 01 '22 at 07:53
  • @ArayKarjauv thanks for helping. I have tried to just feed all these data into the same Neural network, but it has no improvements (actually, only getting worse performance at most of the time). There are better algorithms than deep learning for certain input types. In addition, all these data measurements are unavoidably correlated (for example, a optical fiber will receive measurement when a car passes by. If a car passes by, the car will also make sound, and the camera will receive video). – seermer Jun 02 '22 at 08:13

0 Answers0