Sunday 8 June 2014


Learning Along with Others




ROBERT GOLDSTONE
Psychological and Brain Sciences
Indiana University

VIDEO



Overview: We have developed internet-enabled experimental platforms to explore group patterns that emerge when people can see and imitate the solutions, innovations, and choices of their peers over several rounds.  Experiments and simulations show that there is a systematic relation between the difficulty of a problem search space and the optimal social network for transmitting solutions. With more complex search spaces, people imitate: prevalent options, options that become increasingly prevalent, high-scoring options, solutions similar to one’s own solution, and during the early stages of an extended search process.  Historical records of baby names show that naming choices are influenced by both the frequency of a name, and increasingly by its “momentum” in the recent past.


READINGS:
    Goldstone, R. L., Wisdom, T. N., Roberts, M. E., & Frey, S. (2013). Learning along with others. Psychology of Learning and Motivation, 58, 1-45. 
    Wisdom, T. N., Song, X., & Goldstone, R. L. (2013). 
Social Learning Strategies in Networked GroupsCognitive science37(8), 1383-1425. 
    Theiner, G., Allen, C., & Goldstone, R. L. (2010). 
Recognizing group cognitionCognitive Systems Research11(4), 378-395.
    Frey, S., & Goldstone, R. L. (2013). Cyclic game dynamics driven by iterated reasoningPLoS One, 8(2)
    Roberts, M. E., & Goldstone, R. L. (2011).  Adaptive Group Coordination and Role Differentiation.  PLoS One, 6, 1-8.
    Gureckis, T. M., & Goldstone, R. L. (2009). How you named your child: Understanding the relationship between individual decision-making and collective outcomesTopics in Cognitive Science, 1, 651-674. 


30 comments:

  1. What about stupidity of crown? Sometime imitation is going too far and people follow each other without thinking, it is the case in protest, for example, when someone breaks something there could be a wave of rioters following. It is like people leave there own mind to adopt a crown mind.

    ReplyDelete
  2. Les stratégies pour l'apprentissage social sont très intéressantes.

    Avec l'ensemble des stratégies d'apprentissage social (quand, qui, etc.), est-ce qu'il existe un ordonnancement des stratégies en ordre d'importance?

    Ex. «Copy the majority» et «Copy better individuals».
    Entre ces deux stratégies, la meilleure stratégie a adopté dépend énormément du contexte. Est-ce qu'il y a des recherches qui ont été fait pour faire le lien entre les différents contextes et la stratégie d'apprentissage social a adopté selon le contexte?

    ReplyDelete
    Replies
    1. Translation:

      Is there a hierarchy of strategies, and is optimality context-dependent?

      Delete
  3. I find it interesting that individuals do better when surrounded by imitators. Doesn't there need to be an initial innovator for whom the imitators to copy though? I understand that good solutions are preserved in this manner, but I don't see how people will find better solutions without some sort of innovation. It seems like a "safe bet" philosophy.

    ReplyDelete
    Replies
    1. Yes, if our participants had been surrounded by PURE imitators, they would not have done well. Our result, that being surrounded by people who imitate more, is conditional on the overall natural rate of imitation of our participants ranging only from 0% to 15%. If our paradigm was modified to increase imitation far more than what we observed, there would probably be a level of imitation for which being surrounded by imitators was no longer positive.

      Delete
    2. I think we should be cautious when using the terms imitators and innovators from Rob's experiments to discuss research practices. In Rob's experiment there is no way to have a breakthrough. For example, you can't assemble the creatures in someway that doubles or triples your points.

      If we look at scientific research, on the other hand, there are some innovative breakthroughs that could never happen from imitation. Take the development of the polymerase chain reaction (PCR) for example. Before hand, it took ages to replicate DNA, afterwards we could make millions of copies of a DNA strand within hours. There is no such potential for such a huge improvement in the creature experiment. It is very interesting research that we can learn a lot from when taken in context.

      Delete
  4. Experiment presented show that the more people imitate, the better they do; The more they innovate, the worse they do. I was wondering, to what extent these results are influenced by the setup of the experiment: as I understood, the reward function is well defined, or 'known' in the experiment and stable (quite complicated, but anyway). It would be interesting to see how results change when you start to change the reward function - in other words, introduce a "changing environment".

    ReplyDelete
    Replies
    1. That's a promising idea to pursue. In general, in real ecosystems, the reward function is not extrinsic to the system, but is endogenous. For example, a sparrow survives to breed if it isn't eaten by an eagle, and the eagle survives to breed if it eats things like sparrows. Each animal is providing the reward function for the other. Looking at innovation/imitation in an intellectual ecosystem would provide a way to avoid having to externally stipulate a reward function.

      Delete
  5. Les résultats de l’analyse du jeu ne sont-ils pas particuliers à ce jeu ou du moins à ce type de jeu? Peut-on généraliser et affirmer qu’il vaut mieux imiter qu’innover? Il me semble que cela dépende des situations.

    ReplyDelete
    Replies
    1. Translation

      Does the your team-building case generalize?

      Delete
    2. Je pensais au construction d'équipe de sport dans le monde réel. Je me demandais si dans ces cas aussi il valait mieux imiter qu'innover?

      Delete
    3. Translation:

      I was thinking of team-building in sports.

      Delete
  6. When it comes to social media and blogs and information sharing platforms, I think there might be a influence of the fact that "imitation" is just a click away ("share", "retweet", "reblog"), whereascreating new content implies an additional effort. Wouldn't this also deviate social tendencies towards imitation instead of innovation?

    ReplyDelete
  7. I think that the group average is not a good criterion to assess the effectiveness of the innovation. I would expect the innovators to have a few very successful individuals and many failing individuals. If the rest of the group copy the successful individuals, then it is the innovation that is responsible for the higher score of the group, through imitation.

    ReplyDelete
    Replies
    1. Interesting point. I agree that sometimes, there is one huge innovation (like Bach's Well Tempered Clavier) that others can simply replay and enjoy, and it's hard to imagine improving upon it by tweaking it. But in many other situations, like getting better and better computer chips, machine learning algorithms, or salsa recipes, no individual can fully cover a region of a high dimensional problem space on their own, and so the group benefits from a process of recruiting others to a promising area.

      Delete
  8. 1. Imitation is good for analogue tasks, but how many categories in a dictionary or encyclopedia — let alone sentences expressing propositions — do you think you can convey by imitation? Language is one-trial learning too…

    2. Do you think your patterns are representative of the evolution of Pareto (rich get richer, poor get poorer) effects in real time?

    3. Imitation, like language (and open access) may be a win-win solution for open rather than closed exchange. Or it might not…

    4. Is the team-building task a representative model of cognitive problems in general? Big ones?

    5. Question of scale: Do you think Einstein was an “imitator” or an “innovator”? Or soemthing else? (Incremental innovation seems too trivial to cover it. Maybe insight?)

    ReplyDelete
    Replies
    1. On 5): I would say that Einstein is part of an magnificent chain of researchers both imitating and innovating. He was clearly influenced by Galileo, Mach, Poincare, and Minkowski. In the spirit of "Bad artists imitate, Good artists steal," Einstein fully internalized his influences, and mixed them together in truly insightful ways. I'm betting that Einstein would not have been able to do his work if he had been born in 1700. Needless to say, the kind of "innovations" required of our experiment participants pale in comparison to what Einstein did, and did not require any insight. That's too bad because it could make one feel like imitation doesn't ever involve insight or model-building, which it does.

      Delete
  9. I found Professor Goldstone’s experiments and results very interesting. I am a little worried about how these results generalize to the context of scholarly research. How can we promote innovation or offset its costs in a research context, if penalties are associated with innovation?

    ReplyDelete
    Replies
    1. I think one of the issues is that Prof. Goldstone's talk focused on individual benefits. Academic research is a problem where we'd want to optimize for common output. This said, if innovation leads to suboptimal solutions for individuals, it might well be the case as well for some non-aggregative group activities as well.

      Delete
  10. Thank you for this wonderful talk,
    Imitation => reach and help to keeps a best solution, but you said that in genetic algorithms is not the case, it is more relevant to keeps agents blind in order to get more diversification. Can you give me some more explanation on this!!!

    ReplyDelete
    Replies
    1. I was talking about Kennedy's work with Particle Swarm Optimization:
      http://repositorium.sdum.uminho.pt/bitstream/1822/2291/1/wcci2002.pdf
      He is talking about a multi-agent system that is searching for optimal solutions, and he finds, like we did empirically with humans, that for complicated problems, it is best to restrict how influenced agents are by each other. If agents have complete knowledge about each other, the group converges on decent but not optimal solutions. Likewise, David Goldberg has described a problem with genetic algorithms - all of the candidate solutions begin looking too much alike. The full space isn't being searched anymore. He discussed solutions to this problem like credit sharing (if you have a good solution but others do as well, you have to share the reward among everybody with the similar solution) and similarity-based breeding (agents selectively breed with other solutions that are similar to their own, allowing different solution types to be preserved within the population rather than converging on a population-level mean).

      Delete
  11. I think that in case of social setting the concept of 'fitness landscape' is not very useful, because actions of participants of the system change the landscape itself. This phenomenon is usually referred to as "reflexivity" or coupling between environment and agents: agents make decisions based on their success/failure in the environment, but environment changes as influenced by these actions and that again influences actions/decisions of agents. I am interested in experiments which do not disassemble this relationship (if there are any) and I believe that these may be more relevant for 'real social situations', like scientific inquiry.

    ReplyDelete
  12. I really enjoyed this presentation. I can imagine situations in which the line between innovation and imitation is not as clear cut as in Prof. Golstone's experiment. For example, someone could copy a technique which is used to solve a problem, but then apply this technique to a different sort of problem, and if this was successful, it would constitute an innovation.

    ReplyDelete
    Replies
    1. Yes, this is very common, and I think this kind of in-between case is what causes the advantage to John when John is surrounded by imitators. The imitators are not simply imitating, but are also tweaking John's solution. Your case of tweaking by adapting/generalization a solution is particularly important because it emphasizes that imitation should not be limited to exact, surface-level copying.

      Delete
  13. I found the graphic depicting links between liberal and conservative blogs very interesting. I'm curious to what level this has influenced the recent polarization of political views in North America. If people are discussing their political views online they may only receive praise because they aren't entering the opposition's online territory. I wonder how many people are explicitly aware of this and post very diplomatically written posts on the oppositions blogs, and include links to blogs supporting their opinion. I think its important to have links to various opinions to develop balanced arguments.

    ReplyDelete
  14. Dear Robert, Thank you very much for sharing your very interesting works ! I have one question : Do you think that Wikipedia contributors imitate style of each others and style of other articles to build an (new) article? How characterize and model this kind of imitation in the process of contribution? I would like to think that a computational model of this genesis would improve models for knowledge extraction and ontology learning in Wikipedia.

    ReplyDelete
    Replies
    1. Answer from Robert:

      I haven’t looked into Wikipedia imitation per se, but there’s good evidence out there that people do imitate the language patterns of those around them. For example, in a phenomenon that Bock (1986) calls “syntactic persistence,” people tend to repeat syntactic structures that they have just been exposed to when they are producing their own sentences. This almost certainly goes on among Wikipedia editors too. One felicitous side-effect of this imitation is that Wikipedia pages probably end up being more coherent and sound as though they were “cut out of the same cloth” than they would have if people were not imitating each others’ style. When all goes well, the fact that people are smart about imitating each other means that the works they collectively create end up looking more planned and consistent than you might think given that they are being created by swarms of people.

      Delete
  15. A few people picked up on the idea that some of the results showing imitation as an optimal strategy might be functions of parameters of the problem and of the social environment. One might think that if it's the case, there would be a market for individuals to develop heuristics to recognize the types of problems/environment space they're in, in order to adopt an appropriate strategy. As such, the fact that some people do imitate in the pokemon game could just be a evidence that there are such problem/environments – otherwise, we'd all have already adopted the imitation game.

    ReplyDelete
  16. I think this notion of imitation/remixing has interesting implications for open access and intellectual property. When we acknowledge that problem solving and cultural evolution rely an exchange of ideas, it is readily apparent how closed access, strict copyright laws, and even things like DRM prohibit human development on both an individual and collective level.

    ReplyDelete