Decisions based on perceptual evidence are ubiquitous in human every-day behavior: is that coffee I smell? Was that my friend’s face I just saw in the crowd? Due to ambiguity and noise, such decisions are fundamentally probabilistic and therefore complex. Nonetheless, we make them with remarkable ease and efficiency. This is surprising, given that their combinatorial nature makes following optimal strategies for such decisions intractable. For example, ideally, we would need to consider millions to billions of possible odor combinations just to decide which small number of a large repertoire of possible odors are likely present at breakfast. Similar combinatorial growth of possibilities plagues almost all every-day perceptual choices. This makes the computations underlying such ideal decisions fundamentally unmanageable, such that they need to be approximated. The general question I want to address is which types of approximations our nervous system uses.
An important component of perceptual choices is to accumulate evidence across multiple sources and over time, however brief. By now a large body of work by myself and other has demonstrated that humans and other animals accumulate such evidence in various situations in a seemingly near-optimal manner. Given that the underlying computations ought to be intractable, how is such near-optimal performance possible? The first reason is that in the tested circumstances, the tasks didn’t yet reach the level of complexity that made them intractable. The second reason is that, usually, the perceptual evidence was not well controlled, such that sub-optimal computations could be mistaken for less evidence provided to the decision maker. Once we controlled for this confound it became apparent that approximate computations are already at play for such tasks. These approximations might thus be general properties of our decision-making apparatus. Hence, determining their nature provides first steps towards identifying how our nervous system deals with intractable computations in general.
My research program will focus on the following questions. 1) How much do approximate computations contribute to overall information loss in decision-making? 2) What is the nature of these approximations? 3) To which degree can they be motivated by computational and neurobiological constraints? These questions will be addressed from a theoretical perspective at the level of information coding in neural populations, and in collaboration with others through experiments with humans and animals. In contrast to identifying simply decision-making heuristics from bottom-up, the approach uses optimal, but possibly intractable strategies as a top-down starting point for approximations, with the potential to generalize beyond the specifics of the considered tasks. Thus, the research program would not only give us insight about the algorithms underlying particular decisions based on perceptual evidence, but might additionally yield approximation motives that the nervous system uses in general.