What is it called to classify easy inputs before hard ones? For classifying a lot of inputs, it may be useful to handle the more unambiguous cases first and learn from them before tackling the harder ones. This is certainly how I personally grade student work; the idea must be ancient.
What is it called to classify instances from easy to hard? I thought it would be a kind of reinforcement learning, but I'm having trouble framing this basic idea in the language of RL, as a balance between exploitation and exploration.
 A: The only name I have heard for this principle comes from the legal maxim that "Hard cases make bad law".  This is a very well-known principle of legal analysis that has been popularly expressed in case law for almost two-hundred years.  The reasoning behind this principle is that one is able to induce rules (i.e., classify) more effectively by first looking at simple unambiguous cases, and inducing appropriate rules from these --- the attempt to induce rules from difficult cases is likely to produce "bad" rules.
Although this is a legal maxim, the underlying reasoning for the principle is effectively a principle of inductive reasoning, so it does indeed pertain to statistical inference and classification.  (So you have come to the right place to ask your question --- and as a handy coincidence, your respondent is also a trained lawyer, thus recognising the applicable legal maxim.)  It is probably possible to state the underlying principle you have put forward in the language of statistical inference and inductive reasoning, but I have not heard a specific name for it in this context.
There is no specific name for this principle in a statistical context, since it is just an inherent aspect of the fact that classifications of simple cases are more accurate than classifications of ambiguous/hard cases.  In a statistical context, there is no temporal rule saying that you classify the easy cases before the hard cases, and things tend instead to be done simultaneously, through a model that looks at the totality of all evidence for all casees.  Nevertheless, it is true that more "training" information in a model comes from the simple cases, so then general thrust of what you are talking about is present in a statistical context.
