Arguing Machines: Tesla Autopilot vs Neural Network

[Lex_MIT]

Here's our video of the arguing machines concept applied to Tesla Autopilot "arguing" with an end-to-end neural network. Disagreement is used to predict challenging situations. It's a simple but powerful method for human supervision of black box AI systems:

Paper is available here: https://arxiv.org/pdf/1710.04459.pdf

Some more details on website: Arguing Machines: Human Supervision of Black Box AI Systems | MIT Human-Centered AI

 

[fsl2019]

Hi Lex,

may be a stupid question: In stand still with a static scene (more or less) the network still shows a certain disagreement, where as - obviously - there is no need to steer left or right. Why is that? Is the output of the network in the visualization over exaggerated? Or is that the result e.g. of freespace detection which is often a little bit scraggly? At least what I have seen the freespace detection with NN is not that stable.

Best regards,
Frank