The Use of Public Opinions to Pursue Collective Actions

The Use of Public Opinions to Pursue Collective Actions

When two heads are better than one …
The problem of optimally aggregating diverse opinions together to achieve the best result available is hardly new. The Marquis de Condorcet, 1785, and Francis Galton, 1907, were the first to show how crowds could outperform single individuals as well as experts. More recently prediction markets and Web companies like UNU® have inherited this legacy by finding effective ways to aggregate opinions to forecast successfully future events. Most of the success of these systems can be attributed to one simple mechanism. If we assume people are independent in their judgements, the errors of individuals will be uncorrelated from each other. In other words, some people will over-estimate the true answer, while some others will underestimate it. Averaging together multiple opinions will give however a better estimate because uncorrelated noise will average out, leaving the signal (i.e. the true value) emerge.

…and when two cooks can spoil the broth
So why haven’t we found a better way to make smart collective choices? How is it possible that we often observe crowds getting really wrong answers? Think for example about market bubbles, racism and conformity. The reason behind these phenomena is that most of the assumption of Condorcet and Galton’s theorems are often unmet in reality. People are often not independent in their opinions, we browse the same websites, get our news from the same outlets, we influence each other in unpredictable ways. A rumour might be amplified exponentially or important information could remain not listened to, without any clear understanding of why this is the case. This is why almost nobody among pundits and scientific experts had predicted the outcome of recent political votes. People’s opinions are very difficult to forecast because they are not only based on facts, but also on others’ opinions – which are interconnected by an intricate network of mutual dependencies. Can our scientific knowledge of opinion aggregation help us in avoiding such large-scale social cascades?

Opinion formation and collective action in small groups
Although we don’t yet know the precise mechanisms underlying opinion formation at large-scale, our knowledge is remarkably accurate when it comes to small groups of interacting people. In 2010, Bahrami and colleagues from the University College London showed that when people interact with each other they outperform even the best member of the group. But this effect is found only when people share their confidence in their judgements. Confidence plays a very important role in opinion aggregation as it represents the (inverse) uncertainty of the evidence supporting an opinion. By sharing confidence levels, the group is thus able to select, every time, the opinion supported by the strongest evidence. The following year, Lorenz and colleagues found that knowledge about estimates of others (but not their confidence!) narrows the diversity of opinions to such an extent that it undermines the wisdom of crowd.

In my PhD research I showed that confidence sharing has interesting unexpected by-products. First, people use their own opinions to understand how reliable the information they receive is when no other source of objective feedback is available. This leads people who share the same biases to agree with each other more often, thus reinforcing their views even further. Second, I show that dynamically interacting with others (think about discussing your political views in real life with a friend) greatly differs from simply receiving the same information but without the possibility to interactively discuss (think about reading the same information on the newspaper). Interaction leads to confidence escalation and can potentially produce big errors done with high confidence. This is interesting if we consider that the very same amount of evidence is available in the two conditions.

Opinion formation and collective action in large groups
Researchers have still to link this micro-scale knowledge to macro-scale interventions. Importantly two issues need to be addressed:

  1. How does opinion aggregation scales up with group size?
  2. How does opinion aggregation scales up with problem complexity?

Addressing the first point is relatively easy. In theory, when group size increases, confidence should weigh less and a majority rule like behaviour should be more indicative of the group’s final choice. Although experimental evidence is still needed, the emergence and quick take off of massive online experimentation (e.g. Amazon Mechanical Turk) might soon bridge the existing gap between theory and experimental evidence.

Addressing the second issue might result more challenging. Most experiments to date have involved simple decisions. A simple decision can be defined as a decision where people have to find the optimal answer along one decision variable. Estimating how many people are in a bar you are walking in might be difficult but it’s a simple decision because it involves only one variable (the number of people) that is not dependent on other variables, like the weather or what the bartender had for dinner. Most decisions in real life – think about buying a house or making a new policy – however are complex, meaning that choosing the right answer involves taking into consideration a number of factors and circumstantial evidence at the same time. For example, when buying a house we cannot rely on the price of the house to know whether it is worth the purchase. This has to be considered while taking into account also the size, the position, the neighbourhood and so on. Few studies so far have explored the realm of group decisions in complex decision space. One notable example is the work of Ashish Goel and David T. Lee at Stanford University. These researchers have produced mathematical simulations to find optimal ways to aggregate opinions about complex issues (e.g. making a new policy). They show that a simple majority rule might not be the best strategy available to reach the best outcome. More theoretical and empirical work is however still needed.

Conclusions
Research on opinion aggregation has fascinated researchers and philosophers since Plato’s Republic. However rigorous theoretical work has only recently been coupled with mathematical modelling and experimental evidence. This is a sweet spot in history, where behavioural insights are gaining greater recognition from regulators and governments. The explosion of mobile technologies, Internet data-collection and artificial intelligence allows to speed-up social-science investigations. Quickly understanding these challenges offers a unique opportunity to turn a centuries-old problem into a new millennium opportunity.

Brain mechanisms underlying the brief maintenance of seen and unseen sensory information

I am glad to say that my work with Jean-Remi King and Stan Dehaene has been recently accepted for publication in the prestigious journal Neuron. You can find a preliminary version here.

Recent studies of “unconscious working memory” challenge the notion that only visible stimuli can be actively maintained over time. In the present study, we investigated the neural dynamics underlying the brief maintenance of subjectively invisible stimuli, using machine learning and magnetoencephalography. Subjects were presented with a masked Gabor patch whose angle had to be briefly memorized. We show that the stimulus is  encoded in early brain activity independently of its visibility, and that the maintenance of its presence and orientation can be decoded throughout the retention period, even in the invisible condition. Source and temporal generalization analyses revealed that perceptual maintenance depends on a deep hierarchical network ranging from early visual cortex to temporal, parietal and frontal cortices. Importantly, the representations coded in the late processing stages of this network specifically predict subjective reports. These results challenge several predictions of consciousness theories and suggest that unseen information can be briefly maintained within the higher processing stages of visual perception.

Highlights:

  • The link between working memory and visual awareness has recently been challenged
  • We here study the mechanism of unconscious maintenance with MEG & machine learning
  • Unseen stimuli can be partially and maintained within high cortical assemblies
  • We show how to revise awareness theories to account for the maintenance of unseen stimuli

An Opinion Space to represent Social Interactions

Following my recent presentation for the Experimental Psychology Society meeting in Oxford, I will write about how I came up to conceive the Opinion Space and how I use it now to represent social interactions.

The idea came to me during a project using support vector machines, algorithms in machine learning that represent data points (called instances) in a multi-dimensional geometric space. Separation of categories and learning in this multivariate space is easier than along each individual dimension constituting the space. These sort of machine learning algorithms are great for representing complex data. The signal of different sensors along the scalp, if taken alone, is very weakly indicative of a certain brain state. Taking multiple sensors into account at the same time however (hence “multivariate“) allows us to discern patterns that are not found otherwise.

Social interaction is similar in nature to the problems studied in neuroscience, in respect to its complexity and non-linearity. When two people interact (for example when you talk to friend), they affect each other in very unpredictable ways and with no clear direction of cause and effect. The exchange opinions, views and believes. Each of the individuals in the interaction affects the other (either willingly or not) but is at the same time influenced by the other. By representing social interaction as movements along a higher dimensional space – the Opinion Space -we can understand better the mechanisms, describe better the phenomena and predict better the behaviour of human interaction.

How do we construct an Opinion Space?
Each dimension of the space (feature) is one person’s belief or opinion about a certain variable of interest (e.g. is it going to rain tomorrow? Is the restaurant to the left or to the right going to be better?) or decision (e.g. shall I take my umbrella with me? Shall I try the restaurant on the right-hand side of the road or the one on the left-hand side?). In my research I typically measure the confidence associated with the variable of interest to gauge the strength of the participant’s opinion.

We can now construct from these orthogonal axes a Cartesian space, that we have called Opinion Space, where the information state of the group is represented as a single point along this space. This full Opinion Space is characterised by two agreement quadrants and two disagreement quadrants. We can simplify things to reduce this full space to a more parsimonious version that gets rid of subject’s identity (Yellow and Blue in the video above) and the choice identity (left and right in the video above). We now care only about whose opinion was supported by the strongest confidence (x-axis)? How’s the other person relate to this opinion (y-axis)? Does it agree or disagree? With what level of confidence?

We can now represent the state of the group’s opinion at each moment in time as a point along the space. As soon as one of the two participants change their mind or their confidence as a function of the social interaction, the group’s state will shift to a different point along the Space. We can now track the group’s opinion state as the the trajectory along the Opinion Space.

This method is incredibly useful to compare different social contexts or communication systems, as it allows to quickly visualise the dynamics of opinion formation and social influence. It is also effective to predict the future state of the group give the past and present trajectory. Finally, a simple expansion to more than two dimensions can be used to represent groups composed by more than two members.

The influence of experts on crowds success and diversity

Inspired by the wonderful talks I am listening to in these days at ICCSS2016 and by the recent outcome of the UK referendum, I was wondering how the opinions of experts and news can influence the population.

We know the well established Wisdom of Crowds effect: the average opinion is better because uncorrelated noise (how wrong the opinions are) averages out over large numbers, thus enhancing the signal.

We also know that some people are better than others. We call these people experts. Experts’ opinion is better than the average individuals because it is less variable (more clustered) and closer to the true signal.

Question: what effects do news have when they broadcast experts opinions to the whole population?

  • Does the average error of the population reduce after knowing the expert’s opinion?
  • Does the average error of the population increase after knowing the expert’s opinion?

Simulation
I created a Matlab simulation (that you can download here) that tries to answer this question. You can tweak different parameters like:

  • How gullible is the population, that is how much is swayed by the expert.
  • How many people there are in the population.
  • How closer to the true value the expert’s opinion is compared to the population opinion. This is the ratio between the variance of the expert’s error and the variance of the population error (in the graph this is called “expert’s error”, sorry for the approximation)
  • The number of observations that we average across

Results
Below is the result for a population of 100 people. The colour represents the improvement in accuracy (that is the distance from the true value) of the population average from before to after the expert’s opinion is broadcast.

improvement_expert_sim
Improvement in average population accuracy

The contour line shows the areas of the parameter space where the expert improves the average opinion. Warmer colours indicate better improvement. The blue colour means bad news. I was surprised by two things:

  • the contrast between the tiny area of improvement (on the left of the contour line) and the huge area to the right.
  • the magnitude of the improvement (a small improvement) compared to the magnitude of the decrease in performance (a disaster).

Running the simulation will also output another image showing the decrease in diversity of the population opinion. If you don’t think diversity if important check this out.

Conclusions? No matter how an expert is accurate, there always will be some residual errors in their judgement. Broadcasting that opinion to the whole population has the effect of biasing it instead of helping it. The effect seems to be irrelevant at best, but disastrous in the worst case scenario, that is an expert that is not so much of an expert.

What do you think? Please leave me your feedback!

2nd Annual International Conference on Computational Social Science

This week the second Annual International Conference on Computational Social Science will kick off. I have been invited to give a presentation of my work during one of the parallel sessions. This is the first time I give an oral presentation to an international conference and the first time I present my results on my work on representing information manipulation during social interaction through the Opinion Space.

If you happen to be in Evanston, IL, on Friday 24th of June I hope you will ask many interesting and challenging questions!