Sensitivity to shared information in social learning
In July we’re reading Whalen et al. (2017): Sensitivity to shared information in social learning.
In the study, three experiments are presented with regard to the source of informants’ knowledge in a situation of probabilistic reasoning for the “learner”. More precisely, participants are confronted with a (classic) setup of informants who have drawn coloured balls from one of two urns with known distributions of colours, and have to judge the probability of identifying the correct urn. The results of these experiments are compared to predictions derived from different Bayesian models of rational decisions. The authors conclude that people are sensitive to the amount of information provided by the informants, i.e. whether it derived from independent or dependent sources (“shared information”).
Tell us your thoughts in the comments below.
Thomas Müller 1 August 2017 (16:50)
Sensitivity to information
I take the point that participants are sensitive to the information provided in situations of uncertainty, be it social or asocial information, and that the results of the models show that this sensitivity can be closely predicted by a probability matching account. This is a nice refinement to classical results about conformity biases and social learning, further disentangling the rather oblique concept of the "majority": What are the actual mechanisms involved in its influence, as it cannot have a proximate effect on behaviour?
Still, I also have two worries with regard to the study: First, and this is a common worry I get with experiments employing Mechanical Turk, it seems that a great number of participants in experiment 2 did not pay much attention to the task at hand; the very simple memory check question employed in that part of the study led to the authors dropping ~10% of their participant pool. As a similar check is not reported in either experiment 1 or 3, I have to assume they didn't use it in those cases. Especially in conditions where performance is found to be close to randomness (one case in experiment 1 and two cases in experiment 3), the results then could be due to participants choosing their urn without giving much care or thought to it.
This leads me to the second worry: In a weird way, the comparisons to the Bayesian models also demonstrate that humans are precisely not rational learners as specified during modelling; many of the error bars escape the predicted values, so even if the general patterns tested by pairwise comparisons hold up, participants are not acting in a purely rational way. Likewise, many of the comparisons show no effect where a difference would have been predicted, and in one case an effect is found with no difference being predicted. All in all, I'd still agree the data is "well captured by a rational model of social learning", but also add that there definitely are also many other factors influencing participants' behaviour in this case.
Olivier Morin 18 August 2017 (15:15)
The demise of simple social learning heuristics
This excellent paper interest me chiefly for what it says of recent developments in the study of information exchanges. Three assumptions that used to loom large, at least among cultural evolutionists, are being gradually dropped (or so it seems to me).
- First, as noted by Whalen et al. (bottom of p. 3), "social learning" is no longer considered an alternative to "individual learning." Agents do not need to choose between two modes of decision making, one purely individual, the other purely social. Most human decisions integrate private information with social information. To put it another way, there is no such thing as unalloyed social learning.
- As a result, learning from others must involve fairly sophisticated integration mechanisms. Rough-and-ready rules of thumb that effectively ignore private information ("copy the majority"; "follow the prestigious") won't do. Social learning certainly makes us smart, but not because it is a "simple heuristic." It cannot, at any rate, be simpler than unaided individual learning.
- Accepting these three points leads us to an understanding of social decisions that must be much more congenial to rational choice theory than cultural evolution research used to be. It can no longer be claimed that social learning is a nail in the coffin of Homo Economicus.
Some will say that these developments are nothing fundamentally new, just a natural growth by complexification of the simplified models we start with. My impression, for what it's worth, is that cultural evolution research used to be fairly tightly dependent on a view of social learning as a fast-track way to bypass difficult decisions, one that effectively short cuts more rational mechanisms.