[in construction, links to github and to the lesswrong post will be added when released]

Will Alexplainable construct a system that:

• Can recognize one class (that we can choose as we want) of ImageNet with at least 70% true-positive detection and 70% true-negative detection on all classes

• The system can be explained in terms of how it is mapped out and what part does what

• The system does not use more than 3 layers, ignoring maxpooling and other non-complex types

• No more than 5 hours of setting-up is allowed. The system has to work on its own. There will be clear breakpoints when a human will accept or reject features, possibly explaining why, but a human cannot directly modify the run or intervene.

Once a run is started on a class, that run has to succeed, or a different class has to be chosen