Helping AI to overcome its “blind spots” on road

Jan 28, 2019, 7:14 AM EST
(Source: smoothgroover22/flickr)
(Source: smoothgroover22/flickr)

The memories of fatal crash by a self-driving car in Arizona last year are still raw. What the accident did was to put brakes on overly tall claims and the bid to oversell the technology that needs serious finetuning before practical application.

The researchers at Microsoft and MIT are trying to overcome these “blind spots” with a specially designed training model for autonomous vehicles, reports Engadget. The approach involves AI, reading the behavior of a human driven in a given situation and then comparing it to the action it would have taken.

After analyzing the gaps, the AI alters its response to be as close as to that of a skilled human driver. Research author Ramya Ramakrishnan says, “The model helps autonomous systems better know what they don’t know.” The simulations in the training programs differ from the real-world situations on many aspects, and this approach seeks to fill that gap with the help of humans, notes Slash Gear.