You’re right. I haven’t given it as much thought 💭 as you, but these are the two most common killers — assuming there is sufficient data, the data has signal, and the AI folks know what they are doing.

I have the same problems with my poker DL project, and should address them before just powering through with more modeling instead. Actually… my visualization was a C+ from jump. Asked the network to output extra parameters to guess its opponent’s range and thus “explain” what it is thinking.