So far, I have discussed the production of cinematic time in the context of formal- analytical frameworks to film. As mentioned earlier, there are two significant assumptions that go with these approaches: first, that films are broken down to their constituent units for analysis, meaning frames, shots, sequences; and second, that the viewer submits to a pre- programmed temporal form, for example narrative continuity. In this chapter I question whether and how these assumptions hold under the epistemic displacements brought about by new forms of inductive computing such as deep learning.

Elsewhere I have argued (2012) that computational methods can accommodate radically different theoretical frameworks in moving image studies. Elaborating on this argument, I submit that there are at least two complementary ways to understand the problem of using computers to ‘see’ films for us: what we may call the classical analytical approach, in which the goal is to train computers to perform like critics, and which amounts to designing systems that encode and reproduce existing categories of film analysis. And what we may call generative revisionist approaches, whose goal is to design systems that open new categories and reshape critical practice.

There is no inherent advantage to either approach, nor are they mutually exclusive. They do, however, present distinct advantages and limitations at different junctures. While formalist approaches to film were among the first to embrace the computer as an analytical tool with which to answer existing questions, today, global networked computation is less of a tool and more of a condition under which new questions can be asked and answered.