[IncRnd]

The thing is that humans are generally excellent at these sorts of pattern recognition, but these networks aren't nearly as good. Even in rigorously trained networks that operate surprisingly well, mistakes will appear that if made by a human would be treated as due to carelessness or stupidity.

So, this is going to happen.

[MattGaiser]

> The thing is that humans are generally excellent at these sorts of pattern recognition

Humans with a lot of experience are. Would kids be? I once referred to firefighter as robots as a kid.

[ineedasername]

Our ability for pattern recognition w/ human faces developed over the course of many thousands of years. People in modern fire fighting gear weren't present through that process. A kid thinking a firefighter is a robot is not the same class of problem. Even kids are good at the type of tasks we're talking about.

[MattGaiser]

People have difficulty distinguishing the individual faces of other races.

https://en.wikipedia.org/wiki/Cross-race_effect

So at some level it breaks down for us too.

[ineedasername]

Yes and I think its fair to say that a firefighter in full gear is a reasonable place for it to breakdown, and that does does not otherwise indicate a failing in human capability in this area that should make our failure modes in any way whatsoever at the same level or rate as computer models.

[IncRnd]

The cross-race effect is an instance of the ingroup advantage. However, that can't be extended to say that people will classify blacks as primates.

[IncRnd]

I would say that most kids would not make the same sort of mistake as this network is reported to have made.