Concretely, by exploring the pair-wise and list-wise structures, we impose the relations of generated visual features to be consistent with their counterparts in the semantic word embedding space.

Inpractice, however, atrade-offbetweengenerality (wideapplicability) and specificity (benefitsforspecifictasks) arises. (Figure 1).

Black-Box Optimization (BBO), also referred to as Derivative-free or Zeroth-order Optimization, considers objectivefunctions that arenot knownanalytically and can only beevaluated atvarious inputs,potentiallyatahighcost.

Forexample,theusercouldsolvethe task using actions unavailable to the agent (e.g., the user may have two arms, but a robotic agent mayhaveonlyone) ortheusercould findsuccess examples bysearching theinternet.

Learning-based photometric stereo methods predict the surface normal either in a per-pixel or an all-pixel manner.