The philosopher Albert Camus once said, “Life is the sum of all your choices”. Work using an innovative experimental design in humans and rats shows that many of the errors in those choices come from the senses, not from cognition.
The decisions that humans and animals make define their existence. Some decisions are carefully considered, such as choosing a mate; other decisions must be quick, such as deciding whether a rustle in the bushes is a companion or a predator. In many situations like the latter one, decisions are based on imperfect sensory evidence accumulated over a second or two, so that the brain can average out background 'noise'. Although we do impressively well in making these quick decisions, we make many mistakes. Writing in Science, Brunton et al.1 combined computational modelling with behavioural studies of rat and human decision-making to uncover where the errors in simple decision-making creep in.
Previous work2,3 in humans and primates indicates that there is noise in decision-making, although the source of the noise has been unclear. This 'bottleneck' might be anywhere: in the information from the eyes and ears, in the ability to accumulate the incoming information over time, in arbitrary biases present even before the evidence starts to appear, or in an imperfect strategy that might incorrectly estimate the importance of strong bursts of sensory evidence.
Brunton and colleagues find that errors stem from imperfect sensory evidence, that is, from the initial transfer of these signals to the brain (Fig. 1). So it seems that sensory limitations hold us back from being able to perfectly evaluate incoming evidence.
The authors' experiment was deceptively simple. They presented rats and humans with a precisely controlled but highly variable stimulus — two randomly generated streams of clicks, one originating from the left and the other from the right (Fig. 1). Subjects had to judge whether more of the clicks came from the left or the right. Several aspects of the experimental design allowed the researchers to tease apart where errors came from: the exact pattern and number of clicks varied in each trial; these trial parameters were known by the experimenters and could be exploited in the analysis; and an enormous database of decisions was amassed to increase the statistical power of the results.
The authors ruled out several candidate explanations for why subjects make mistakes. First, they observed that subjects performed equally well on trials that had many different patterns of clicks. This allowed the authors to exclude the possibility that errors were caused by subjects being swayed too strongly by bursts of clicks. Second, the authors evaluated decisions in response to stimuli of differing durations. Brunton et al. could thereby rule out two more sources for subjects' errors: 'muddling' of the active memory owing to the passage of time and 'forgetting' clicks that came early in the trial.
The lack of muddling and forgetting is surprising; although the fundamental timescale of neuronal activity is fast, on the order of 10 milliseconds, it seems that the brain as a whole can avoid losing information over the several-second stimulus duration that defined the longest trials in the study. The long timescale of evidence accumulation points to a network of neurons that collectively keep the memory of early sensory evidence alive4. Noise must therefore enter earlier in the process.
The authors went one step further: they formulated a flexible mathematical model to determine each source of noise quantitatively. Although previous studies5,6 had concluded that the brain can reliably accumulate evidence over time, this study is the first to simultaneously provide a quantitative estimate of the timescale involved in decision-making (at least a few seconds) and of the other sources of noise.
Brunton and co-authors mainly studied rats in this work. The rat is a relative newcomer to the field of perceptual decision-making. But although primates remain indispensable for neuroscience research because of their greater ability to be trained and their evolutionary closeness to humans, rats offer some key advantages. First is higher throughput: rat experiments can be conducted faster than primate ones. Consequently, rodent studies can include a large cohort of subjects, making it possible for experimenters to distinguish average behaviour from idiosyncratic behaviour. Second, a growing battery of molecular tools is becoming available for use in rodents, allowing easier tracing and manipulation of specific brain circuits7.
The present paper opens the door to a number of new directions in the study of decision-making. The ability to pin down the source of noise could reach well beyond the task for which Brunton et al. used it. Whereas their results suggest that most of the wrong choices the subjects made were due to errors early in the sensory pipeline, with different tasks the errors might be driven by other sources of noise. The authors' modelling techniques could be adapted to explore the source of errors in other tasks, such as value-based decision-making8,9.
An appealing possibility is that the sources of noise will be highly task-dependent. For example, memory might form the bottleneck when comparing stimuli over long periods, as when searching for a new apartment. Or perhaps pre-existing biases dominate when the cost of switching choices from one trial to another is high, as in foraging tasks in which subjects must choose whether to keep exploiting their current, but diminishing, resource or to pay a cost to explore another10. It would also be intriguing to compare the magnitude of the sensory noise in multisensory decisions11, in which the presence of two sensory inputs might increase the sensory noise.
In all types of decisions, estimation of noise parameters could be a starting point for investigating neural circuits. Many open questions remain in the field of decision-making, both in terms of the network of neural structures that are required for different kinds of decisions and the microcircuits that support computation within those structures. With the host of tools now available for dissecting neural circuits during decision-making, the coming years could bring major insights into the mechanisms underlying this crucial ability.
About this article
Proceedings of the Human Factors and Ergonomics Society Annual Meeting (2015)