Post by Stephanie Williams
What's the science?
The human decision-making process includes weighing relevant information and uses these weights to select a choice. Currently, there are several models of decision-making that explain how this happens in terms of an "evidence integration" computation. Although informative, these models have primarily focused on decisions lasting on the order of seconds. It is still unclear whether decisions over longer periods of time can be modelled in the same way, or whether fundamental memory limitations would prevent humans from using integration over longer durations. This week in Current Biology, Waskom and colleagues designed a new task that probed evidence integration over longer periods of time..
How did they do it?
Five participants saw a series of patterns on a computer screen. The patterns were shown at varying levels of contrast against a grey screen, such that some patterns barely contrasted against the grey background, while others stood out. The set of 1-5 patterns, presented sequentially, was randomly sampled from either a) a low contrast distribution or b) a high contrast distribution (see figure). Participants had to decide whether, overall, the series of patterns came from the high contrast distribution or the low contrast distribution. There were either shorter (1-4s) or longer (2-8s) unpredictable gaps between the patterns they saw in each series. When making their decision, participants were instructed to think about the average contrast of the patterns they saw. The authors manipulated several parts of the task— the strength of the contrast of the pattern, the number of patterns (1-5), and the length of the gap between each pattern. To perform well at the task, participants should have used all of the patterns presented to them in the series to make a decision.
The authors evaluated each participant’s behavior by looking at how the three aspects of the experiment they manipulated influenced the subject's choices. They investigated each of these aspects in both individual subjects and in the aggregated group. They then fit four different computational models and compared the predictions of the models with their data to infer characteristics of the decision-making process over longer timescales.
What did they find?
The behavioral data showed that participants were able to accurately integrate evidence over periods of time on the order of tens of seconds (ranging from 2.2s to 34s). Participants were sensitive to the strength of the contrast in the patterns they saw, and performed better when they saw a greater number of contrast patterns before having to make a decision. Importantly, participants performed with similar accuracy in both task conditions (long vs. short gaps between stimuli), suggesting that evidence integration is a flexible process that can extend across long timescales. Of the four models the authors examined, a linear integration model best fit the data, suggesting subjects summed the evidence from each pattern they encountered to make their decision. Directly modelling two proposed sources of information loss, ‘memory noise’ and ‘memory leak’ (when information presented earlier is forgotten), showed that neither were present in any appreciable magnitude. The subjects’ data were not perfectly explained by the linear integration model, however. Subjects tended to slightly overvalue stimuli that appeared first in each trial, suggesting that they sometimes struggled to change their mind after forming an initial impression.
What's the impact?
The authors’ findings advance our knowledge about how we combine evidence at timescales corresponding to many real-world decisions. The study shows that people are able to integrate data with minimal information loss over relatively long durations. The findings also pose important questions about the biological mechanisms behind evidence integration during natural decision-making, and suggest current network models may need to be amended.
Waskom et al., Decision Making through Integration of Sensory Evidence at Prolonged Timescales. Current Biology (2018). Access the original scientific publication here.