3 ms·

Computer vision is advancing exponentially. This year the imageNet machine vision challenge winner got 6.7% top-5 classification error. 2013 was 11%, 2012 was 1

Computer vision is advancing exponentially. This year the imageNet machine vision challenge winner got 6.7% top-5 classification error. 2013 was 11%, 2012 was 15%. And since there is an upper limit of 0% and each percentage point is exponentially harder than the last, this is actually significantly better than it seems. That's about as good as humans are estimated to be at this task, and humans require much more time to search through reference images. The future of robotics is not manually programming robots with every move they should make, using handcrafted features to do vision,like your towel robot. It's letting the robots learn how to do tasks by themselves.

dragontamer 12y ago
''' The future of robotics is not manually programming robots with every move they should make, using handcrafted features to do vision,like your towel robot. It's letting the robots learn how to do tasks by themselves. ''' That's still programming. If you've ever touched "Logic Programming Languages" like Prologue, you learn that you leave an enormous amount of reasoning to machines today. For the more mainstream programmer: it is a brutal amount of AI that is used to calculate "auto" specifications and template parameters in C++. But in either case, the AI may be significantly faster or accurate today... but you still require a programmer that understands how to use the AI and give it direction. True, not every move needs to be programmed anymore. But neither are we programming in Assembly language. We've got huge advances in compiler technology (which really are stemmed from the automated reasoning groups in AI). At the end of the day however, the AI's purpose is defined by the programmer.
- Houshalter 12y ago
  Yes of course the thing will have to be programmed to some degree. The point is that it's within the realm of possibility to use AI to do computer vision and robot control, rather than having some programmer trying to figure out every possible edge case of a messy real world problem. Instead the programmer designs the reward function and collects data and sets the hyper parameters, then the AI figures out how to solve the hard parts of the task itself.

dragontamer 12y ago

''' The future of robotics is not manually programming robots with every move they should make, using handcrafted features to do vision,like your towel robot. It's letting the robots learn how to do tasks by themselves. ''' That's still programming. If you've ever touched "Logic Programming Languages" like Prologue, you learn that you leave an enormous amount of reasoning to machines today. For the more mainstream programmer: it is a brutal amount of AI that is used to calculate "auto" specifications and template parameters in C++. But in either case, the AI may be significantly faster or accurate today... but you still require a programmer that understands how to use the AI and give it direction. True, not every move needs to be programmed anymore. But neither are we programming in Assembly language. We've got huge advances in compiler technology (which really are stemmed from the automated reasoning groups in AI). At the end of the day however, the AI's purpose is defined by the programmer.

Houshalter 12y ago
Yes of course the thing will have to be programmed to some degree. The point is that it's within the realm of possibility to use AI to do computer vision and robot control, rather than having some programmer trying to figure out every possible edge case of a messy real world problem. Instead the programmer designs the reward function and collects data and sets the hyper parameters, then the AI figures out how to solve the hard parts of the task itself.