Thus far, machines have not done well in scenarios of low stationarity. Those are scenarios where the past is not so good a predictor of the future or where the data manifold is rapidly changing. This occurs for example, when the act of predicting changes what is being predicted. The related notion of Antipredictable Sequences is the most interesting consequence of the oft misconstrued notion of the No free lunch theorem.
Humans are far from the ideal machine for this scenario but they vastly outmatch current algorithms. This is because humans are much better in the kind of generalization that is based off theory building. While not perfect, it does well enough at generalizing from observations (think Newton deriving the laws of gravitation from observations and Brahe's data) to scenarios that are not so close to the data.
This means that for now and in complex environments, simpler models that are adaptable to change online are better. Additionally, their inspectability means they are more easily modified to better match any very changed dynamics.
So your examples of: `speech recognition, face recognition, object classification and detection, speech`
Are of the type where a learner, such as a deep network, whose generalization ability mostly hails from interpolating to fill in missing structure based on data (implying poor out of sample performance) does well when you can give lots of examples which provide good coverage.
Language translation is sort of inbetween (the space is not as smooth but is still fairly stationary at the scale of years) and so we also find that the advantages of deep learning have not been as pronounced and dramatic as has been the case for image and speech.
The digital machine advantage is two. Their memory capacity is outmatched. Second, and related to the first, energy needs are not so pressing a concern.