Inferential Representations

The starting point is a situation from which we want to get certain information, and there is something we want to know about it. In order to access this information, we proceed by first producing a representation of the situation: one that represents not the whole of the situation at hand, but (at least some of) its relevant aspects in a tractable way. In other words, we want a representation that carries enough relevant information about its subject, without adding too much noise, that is, without including elements that might be mistaken as representing something relevant about its subject without actually doing so. Once we have such a representation, we translate our original question about the target situation into a corresponding question about its representation. If the system works properly, then the answer we get to this second question will serve us also to answer the original question, once properly translated.

Notice how in drawing on the representation to get information about the situation, we perform several inferences, and not all on a par. In broad terms, we an identify two different – and equally important – sorts of inferences: inferences of the first sort go from features of the representation to what the representation represents, while inferences of the second sort go from what the representation represents to how the world is. In the first sort of inferences, we go from information about the representation – mostly, about how it is, but also background information about it – into information about the representation’s content, i.e., about how the world is according to the representation. Inferences of this sort constitute what is commonly called the interpretation of the representation.

Once we have performed these inferences and have determined what is represented in the representation; we still need to proceed to inferences of the second sort. Once we know how the world is according to the representation, it is important to determine whether the world is actually as the representation represents it to be. Once again, background information about the representation and how it was created will be essential.

Sometimes, in order to extract from a representation the information we need, it might be enough to just inspect the representation, looking for the relevant information in it. However, other times, the solution might not be so straightforward, and some manipulation of the representation might be required to extract the relevant information. For example, in photo finish situations, it is common practice to draw parallel lines on top of the photographs marking the edge of each racer closest to the finish line. In this process, it might be also necessary to combine the information contained in the representation with background knowledge from the target situation itself. When using a map to navigate a city, for example, it is commonly necessary to match information from the map and information available at the context of use to determine what route to take or even to identify just where one is. In other words, sometimes, there is a going back and forth between representation and target in order to interpret the representation, and thus acquire the desired information.

Most of the times, when we use a representation to perform an inference, we need to proceed through these two sorts of inferences. We need to first interpret the representation, before we can infer something from it about the world. After all, information contained in a representation is useless if it is not extracted from it, and applied to the world. Both elements are fundamental. Different sorts of representations require different sorts or interpretation, and different representations are trusted for different reasons. When we see the light flashing through the numbers inside an elevator, for example, we trust the information it gives us about the elevator’s itinerary for different sort of reasons that why we trust a photograph we took ourselves, or the maps on a subway station. Similarly, how we interpret a text is significantly different to how we interpret a map or how a radiologist interprets an X-Ray image. When talking about images, heterogenousity is the norm.

Notice furthermore that, in very case, the whole process involved in using a representation to perform an inference is rather complex: it involves, not only the generation of the representation of the target situation, but also the interpretation, manipulation and evaluation of the representation. All this to draw an inference that, at least in principle, could have also been achieved working directly on the world. This means that it makes sense to use a representation to draw an inference, only when its use has some advantage, either in perspicuity, certainty, accessibility, etc. over working directly on the subject of such inference. This means that what makes a representation good for a certain inference, is not just its accuracy in representing its target or its reliability in producing valid inferences, but also its usefulness: its tractability, accessibility, clarity, etc. In other words, it must not only be effective in giving us the information we need, but must do so in an efficient way (Giardino 2012, Kulvicki 2010, Blackwell 2008,).