FSD Predictions for 2020

[strangecosmos2]

The above post links to MobilEye that calls it "an end-to-end deep neural network, which may be trained to predict the correct short range path from an input image".

Which post? Don't see the link you're referring to.

The thing I want to distinguish is a fully end-to-end system like Wayve's and a mid-to-mid system like ChauffeurNet and a system like Tesla's that may (or may not) use end-to-mid for one or two or a handful of tasks — like path prediction* — out of ~100 tasks overall.

*As I said above, I'm not sure whether the input for path prediction is raw pixels (which would make it end-to-mid) or computer vision representations like road edges (which would make it mid-to-mid).

 

[EVNow]

Which post? Don't see the link you're referring to.

Unless you blocked the poster - you should be able to see this.

FSD Predictions for 2020

 

[strangecosmos2]

Unless you blocked the poster - you should be able to see this.

FSD Predictions for 2020

Thanks. So, what Mobileye has patented appears to be a path planner that includes both a mid-to-mid imitation learned module and an end-to-mid imitation learned module:

 

[EVNow]

Thanks. So, what Mobileye has patented appears to be a path planner that includes both a mid-to-mid imitation learned module and an end-to-mid imitation learned module:

They shouldn't allow obvious things like this to be patented.

 

[Galinette]

Are you going from memory ? I don't recall him saying theoretically that's the case. Either way - given the challenges with getting end-to-end to work, I suspect weeding out bad drivers is a relatively easy problem.

Even if you filter out bad drivers, you will not have a system which is safer than a human driver with this approach. It will learn how to be not better than a human driver, with reaction time lag, etc... You cannot use any driver input database to train AP. Unless you only use only pro driver input specifically recorded for the purpose and find a way to compensate for reaction time. Or alternatively ask Chuck Norris to train the NN.

 

[Galinette]

They shouldn't allow obvious things like this to be patented.

It’s quite easy to patent stupid things, despite patent office reviews. A patent is just a base for legal fights and in this case, the patent is weak.

 

[Mardak]

Even if you filter out bad drivers, you will not have a system which is safer than a human driver with this approach.

Neural networks are good at filtering out noise and learning the common case, so even with bad drivers, the average and good will likely balance out, but even then, having good data is better. Human drivers generally drive okay and avoid accidents, but people can get distracted or have moments of bad judgement, so a computer driving just at an "average" level but with perfect execution can be better than the average human.

 

[strangecosmos2]

I suspect this problem is soluble with driving because it was soluble with StarCraft: