2
$\begingroup$

If I have a car and a human is driving a car on a daily basis. I create a system to analyze the human routines of driving and implement this system in maybe 10,000 cars. I use that data to train an AI for driving.

  • How much computation I would require to train this system if i am using image processing and human reaction to those images. to elaborate this I mean to say processor or GPU I would require.

A rough estimation would be satisfactory.

$\endgroup$
  • $\begingroup$ Hi and welcome to this community! Please, ask just one question, otherwise, this question should be flagged as too broad. Also, be more specific. What resources do you need for what? $\endgroup$ – nbro Jul 31 at 10:03
  • $\begingroup$ @nbro thanks of you comment. I tried to be more specific and I was just tried to get just how can I achieve this task. and how a person who is working on ai will view this scenario. $\endgroup$ – xitas Jul 31 at 10:18
  • 1
    $\begingroup$ @nbro the two questions seem related (how much computation/to what scale.) I've edited the question header, but the question itself could probably use another edit. $\endgroup$ – DukeZhou Aug 1 at 20:14
  • 1
    $\begingroup$ @DukeZhou Even if they are related, I would say that it is always preferable to ask one question per post! Also, it is not clear what "To what scale it would be impossible to operate in" means exactly. $\endgroup$ – nbro Aug 1 at 20:20
  • 1
    $\begingroup$ @nbro I edited out the second part of the question b/c the first part of the question likely has more utility, in general, in terms of the resource requirements for this type of machine learning. Since we don't have a "resource consumption" tag, I utilized the computational complexity tag. $\endgroup$ – DukeZhou Aug 1 at 20:28
1
$\begingroup$

From an AI perspective the described setup make sense. The human driver but not the computer controls the car which allows to drive the vehicle in the normal traffic. The only modification is, that all the actions are recorded, and maybe the face of the driver as well ... From the computational situation a real time recording of the traffic around the car is a demanding task. Highend equipment is needed which includes Lidar, some high resolution cameras, object detection algorithm and the trajectory tracker. Such a car will be equipped with a huge server rack at the backside and additional power supply is needed too.

This setup describes only the data recording, while the human driver is in control of the steering wheel. If the idea is to convert the rawdata into a self-driving car software many more effort is needed. In the easiest case the corpus of driving examples is analyzed by a team of phd students who are creating an LSTM neural network around the data which is able to control the TORCS simulator. If the project should become more reliable, the vehicle controller have to be grounded with natural language which results into long term planning like “crossroad ahead, drive slower”.

$\endgroup$
1
$\begingroup$

It's called "imitation learning" or "behavior cloning". Basic NVIDIA demo from 2016 was trained on only 72 hours of driving. Model of that scale (it's a small scale) usually require days of training on single GPU. In general this approach doesn't work well beyond demos because of "edge cases" - situation the real driver encounter rarely but which lead to accident and also general instability of cloned driving policy. That's the reason why research in that direction use simulators a lot. This approach didn't make a lot of progress in the last couple of years and mostly went out of fashion in mainstream autonomous car research. Still there is a plenty of academic papers produced in that directions even now. You can google on "imitation learning autonomous car" and "behavior cloning autonomous car"

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.