In machine learning 95% accuracy is not very useful. It means the algorithm will be wrong 1 out of 20 times. So every 20th word in speech recognition will be

In machine learning 95% accuracy is not very useful. It means the algorithm will be wrong 1 out of 20 times. So every 20th word in speech recognition will be incorrect, your medical diagnosis will be wrong for every 20th person, your face will not unlock your phone every 20th time, every 20th person will not be allowed to log into a web site. There are some machine learning applications where 95% is good enough or even great such as web search engines, but for many important applications it is nowhere near good enough and you need something more like 99.9% or better to be actually useful. Think about how often the algorithm can be allowed to make a mistake. If it's never, then don't use machine learning. If it's 1 in a million you need 99.9999% accuracy. 95% looks good in machine learning papers in computer science, but in the real world 95% often means dismal failure. Think of an autonomous killer robot, is mistakenly killing 1 out of 20 people ok? For a self driving car is harming the driver or car (or pedestrian/dog/child/other car) mistakenly even 1 out of 1000 (99.9%) times ok? I think not. One measure of what accuracy you need is how well people do the same task. If you can do better than people most of the time, maybe the algorithm is useful, though you also have to compare exactly how the algorithm fails to the way people fail. If people fail by choosing a fender bender and the algorithm fails by killing someone, the algorithm has to work better than people do.