Coding Subjective Value in Preference-Based Decisions

Post by Elisa Guma

What's the science?

In our daily lives, we are constantly processing information in order to make decisions. Due to capacity limitations, our brain has had to adopt various strategies for information processing. One such strategy is known as efficient coding: the more we are exposed to a certain environmental stimulus, the more precise our neural representation of the stimulus is. Another model of how we process information is Bayesian decoding, which suggests that for optimal processing we combine our representation of environmental stimuli with our prior expectations. Recent work suggests that the brain can adapt to using different strategies based on the environment, however, it is currently unknown whether efficient coding and Bayesian decoding principles are used jointly to generate subjective value preferences, and whether these can explain variability, biases, and confidence in value-based decisions made by humans. This week in Nature Neuroscience, Polania and colleagues aimed to show that choice variability, biases, and confidence in human preference-based decisions can be explained by single value-inference process using modeling and behavioural experiments.

How did they do it?

The authors recruited healthy young volunteers (n=127) to participate in one of four different behavioral experiments in which they were tasked with assigning preference-based value to different food images. Experiments 1 and 2 each consisted of two different rating phases, followed by a decision-making task. In the first rating phase, participants were shown images of 64 food items and asked to provide a rating (on a continuous sliding scale) based on how much they would like to eat that food. During this phase, participants were unaware that there would be a second session, in order to prevent them from trying to memorize their ratings. Items were randomly presented for the second rating phase, in which participants followed the same instructions as the first. Experiment 2 was the same as Experiment 1, except that participants used a 20-point scale for ratings instead of a sliding scale to control for the possibility of anchoring biases. Immediately after the two rating sessions for Experiments 1 and 2, participants were put through a choice task, in which they had to make a preference-based choice between two food items (shown during the rating phases) whose ratings differed by either ~5%, ~10%, ~15%, or ~20%. In Experiment 3, participants were presented the same food items as in Experiment 1, except that half the images were randomly selected to be shown for 900ms, and the other half for 2,600ms to test whether longer exposure would improve accuracy of the encoding. Finally, Experiment 4 was the same was as Experiment 1, except participants also had to provide a confidence rating for each food preference rating. The authors used various modeling techniques to determine how the presentation of an object with a true stimulus value (the ground truth) elicits an internal noisy response (encoding) that is then used by the observer to generate a subjective value estimate (decoding). They used hierarchical logistic mixed-effects regression as well as Bayesian modeling to decipher how the preference-rating variability and value difference affected choice. Predictive accuracy of their models was tested using leave-one-out cross-validation (a machine learning technique).

What did they find?

In Experiment 1, the authors observed that in the choice phase, the greater the difference in value participants assigned to the two food items (during the initial two rating phases), the more consistent the choice between the rated items. Further, the greater the variability in the rating given to the item, the less consistent the choice. In Experiment 2, they observed a similar impact on choice behaviour as in Experiment 1, confirming that the rating procedure (continuous vs. fixed scale) didn’t influence rating variability. In Experiment 3, they found that a longer exposure reduced ‘internal noise’ of individuals’ subjective preferences, decreasing rating variability (these findings were also confirmed theoretically using a mathematical proof). Finally, in Experiment 4, the authors showed that the confidence in an individual's’ rating relates to the rating variability (independent from the actual rating). Qualitative predictions and leave-one-out cross-validation suggested that human preference-based decisions are inferred and employed using both efficient coding and Bayesian decoding, but that the efficient coding model predicted the data best.

What's the impact?

This study found that variability, biases, and confidence in preference-based choices can all be explained by one specific encoding-decoding strategy optimally suited for a limited-capacity system. This work provides a new framework for modeling and predicting preference-based decisions and furthers our understanding of how humans perceive and evaluate their environment in order to guide behaviour. This work can be expanded to other fields such as psychology and economics in order to further understand decision-making strategies accounting for environmental and biological constraints.

Polania et al. Efficient coding of subjective value. Nature Neuroscience (2018). Access the original scientific publication here.