I. Introduction

Why are humans the only species exhibiting exponentially accumulative culture? Language obviously currently facilitates this process, but language is also an example of an accumulated cultural artifact. The development or evolution of language at best may have co-evolved with our cultural acquisition capacity, but probably this capacity must have preceded it.

We know other species regularly exploit socially-transmitted behaviour (Franks and Richardson, 2006; van Schaik et al., 2003; Perry et al., 2003; Galef Jr. and Laland, 2005; de Waal and Johanowicz, 1993; Whiten et al., 1999). Thus the basic capacity of social learning is present in these species, and further has proved adaptive at least in limited forms.

In this article I review the representations underlying social learning. I then propose that for most species the adaptive rate of cultural evolution is bounded by ecological pressures, but that in the case of humans a uniquely rich representational substrate has allowed the coevolution of intricate norms and behaviours. This allows human cultural evolution to find sustainable behavioural strategies.

II. Representations Underlying Social Learning

Learning is easy. Assuming that all you mean by `learning' is changing values inside a representation. Constraining learning so that it does something useful is the hard problem. This was the conclusion of Marler (1991) after he examined the surprising diversity of mechanisms that have evolved to satisfy one relatively simple problem: the transmission of birdsongs between individuals of a species. This has also been the experience of artificial intelligence (Bishop, 2006). The lesson from the current emphasis in machine learning on Bayesian statistics is that the hardest part of learning is establishing a characterisation of the learning space (Wolpert, 1996; Chater et al., 2006). In Bayesian terms, the problem is selecting an appropriate class of models for the learning domain, part of the task of establishing an appropriate set of prior probabilities. These results of the mathematical analysis of the complexity issues of the learning problem may be used to explain the fact that the vast majority of learning in nature is carefully specialised to task (Gallistel et al., 1991; Roper, 1983). The results further suggest that for species that do possess some general learning capacity, the probability of an individual stumbling across a useful piece of knowledge within its lifetime is not necessarily high.

Where an individual agent, either animal or artificial, does have the capacity for general learning, it may very well be in its interest to learn knowledge that has proven useful to other similar agents. This line of reasoning has lead to the recent surge in interest in culture and social learning.

This section begins with a basic taxonomy of social learning. From this I will derive which representations are needed to explain the various components of such learning. This will begin to differentiate the capabilities of different species, and help explain what determines the fidelity or granularity of behaviour replication. This will in turn help return us to the question of what makes human culture different.

Decomposing Social Learning

This is a simplified taxonomy of the forms of generic social learning exhibited in nature. For more elaborate taxonomies and more complete descriptions, see Zentall (2001) or Whiten (2006). In the below descriptions, the model is another agent that already holds and expresses a behaviour being socially learned.

Here I have not addressed the highly specialised, species-specific, evolutionarily `ritualised' forms of learning, such as tandem running in ants or imprinting in hatchling birds. Although in reality, the extent of general-purpose learning is often overestimated (Gallistel et al., 1991; Roper, 1983). Even simple stimulus-response conditioning does not work for all stimuli to all responses. Pigeons can learn to peck for food, but cannot learn to peck to avoid a shock. They can, however, learn to flap their wings to avoid a shock, but not for food (Hineline and Rachlin, 1969). Similarly, rats presented with `bad' water learn different cues for its badness depending on the consequences of drinking it. If drinking leads to shocks, they condition to visual or auditory cues, but if drinking leads to poisoning they learn taste or smell cues (Garcia and Koelling, 1966).

Such examples indicate learning biases in the brain, at the level of associative learning. Also often overlooked are biases provided in terms of motor and perceptual capacities. If an animal cannot perceive something (e.g. colour), it may be because no chance mutation has ever lead to that capacity, but it may also be because colour perception adds no net value to that animal's behaviour repertoire, but might distract it with irrelevant detail.

Primitive Elements of Social Learning

Computer scientists -- including those who build Artificial Life (ALife) and Artificial Intelligence (AI) -- often speak of `primitives`. Primitives are the fundamental components, the building blocks of behaviour. Like atoms, these primitives are themselves constructions (e.g. of neural coding). But any particular discussion requires basic units at some level of abstraction. In ALife or AI the primitives are built in conventional computer code. The intelligence of the system then must express them in reasonable contexts and orders. There is some evidence that brains also work this way, with complex gestures and stimuli being represented (and even generated) by single nerve cells (Perrett et al., 1987; Rizzolatti et al., 2000; Graziano et al., 2002).

In an attempt to understand the representations underlying social learning, I will begin by defining a few primitive elements or actions. Notice that not all of these primitives will necessarily appear in the final theory. Rather, I am starting with a set of primitives I believe underly common theories of social learning, but I will not necessarily support them.

Notice that the preceding definitions of elements specific to social learning necessarily imply a set of representational primitives: contexts, goals, actions, body parts, and coordinates. In addition, we might also assume the presence of several more general-purpose abilities:

Again, I am not proposing that all these capacities are available in all (or even any) agents capable of social learning. I am claiming that these capacities are needed in order to display all the forms of social learning mentioned in the original taxonomy of social learning.

Analysing the Taxonomy

To begin with, there is no social learning without individual learning. In fact, social learning can be seen as a special case of individual learning -- a set of evolved biases for acquiring information by exploiting the knowledge of others (Bryson and Wood, 2005; Wood and Bryson, 2007).

Social facilitation, location enhancement and stimulus enhancement are very little more than individual learning. Location and stimulus enhancement assume context identification, plus either an association with an established behaviour or the individual learning of a new behaviour, either of which occurs as a consequence of being attentive to the location or stimulus. These forms of social learning in no way assume goal, action or body mapping. Social facilitation requires no learning at all, although it may result in learning that increases the probability of associating some context with some known action, in the case where the social facilitation keeps happening in the same context. An example of this case of learning resulting from social facilitation might be the gradual social tuning of the context in which innate warning cries are expressed by vervet monkeys (Seyfarth et al., 1980).

Goal acquisition through emulation might seem as simple as stimulus enhancement, since it might also require the acquisition of a single primitive element, the goal. However, motivations are fundamental to an agent's intelligence, and it is not easy to see how a totally new goal would be incorporated into an agent, with its associated drives and emotions. Goal emulation may be more like operant conditioning. An action or a perceptual context might become identified with a pre-existing drive, and thus become desirable itself. This reduction can be applied to simplify or eliminate goal mapping as a primitive. Goal emulation could be accounted for through action mapping, with the additional recognition or association of the observer's own desire to its perception of the target's action. If the two animals are in a similar state, whether due to shared history (e.g. a troop hasn't eaten yet today) or shared responsiveness to a perceptual context, then the probability of sharing a drive may be high enough for reasonably accurate learning to occur.

At its simplest then, goal emulation might be viewed as the association of a behaviour to a context, where that context is some combination of a perceptual context and an internal drive. Put even more simply, it is socially acquired stimulus and response.

Program-level imitation (or staged emulation) is essentially a set of goal emulations -- or a structured association of contexts to actions. The extent of this structure is much debated. It is tempting to take what appears to be the simplest explanation, and assume that simply associating sufficient perceptual context (perhaps including recent memory of prior events) to action responses will allow you to have an otherwise undifferentiated set of stimulus-response pairs to form the representation for learning a new task. However, this is not what humans or animals appear to do. In extensive experimentation with modelling human learning, Anderson et al. (1997) determined that intelligence driven by sense-action pairs requires specification of a subset of pairs to be active in a particular task context. Even within the task-specific subset, they also require each pair to be associated with a probability for being useful, referred to as a utility value. In my own research, I have found evidence that even this amount of information is not sufficient. Rather than probabilities of success, accurate representations of priority of one task-element over another are needed to guarantee task consummation. We have evidence that this better describes the behaviour of monkeys at least (Bryson and Leong, 2006; Wood et al., 2004), as well as being a useful representation for organising artificial intelligence (Bryson and Stein, 2001; Bryson, 2003).

There is also evidence of neural representations for meta-level task information such as order in a sequence (Tanji, 1996).

Whiten (1998) has reported that chimpanzees not only imitate hierarchical behaviour, but do so more accurately on subsequent trials if the demonstration is repeated. This increase of fidelity -- essentially moving from goal emulation to program-level imitation -- might result from better learning the affordances of the task facilitating lower-cost representation and thus easier learning. Or there may be a social drive to emulate with more care when prompted by a repeated demonstration. However, these increasing-fidelity results have not yet been well-supported through replication, although the hierarchical structure of social task learning has (Whiten et al., 2006).

If there are limits to the number of discrete task steps that can be imitated programmatically, then this indicates that gesture imitation (by which I include vocal-gesture imitation) may require a completely different representation. One could imagine that gestures could be extended sequences of many body or coordinate mappings. However, it is well-established that there is no neurological means by which rapid sequences of action expression can be launched independently each in response to the other (Lashley, 1951; Henson and Burgess, 1997; Davelaar, 2007). In other words, muscle firing cannot be integrated.

I believe the capability for high-fidelity, temporally-accurate gesture imitation may be the key to the puzzle posed in the introduction to this article -- why human culture is different, at least from other primates'. I will explain how and why below. But for now, I will stick to the issue of representations. The relevant data here is that there is no evidence primates other than humans have the representation necessary for gesture or vocal imitation (c.f. Fitch, 2000, for the vocal case in particular).

But if apes are not capable of full gesture imitation, how can they perform “do as I do” tasks? These involve imitating the gestures of a demonstrator (normally human) such as clasping one's self, or jumping up and down (Custance et al., 1995). These sorts of imitation certainly do require some kind of body mapping, and a process of action sequencing. But because chimpanzee and human bodies are similar, it may be that a very low resolution representation of the body configuration at the start and end points of the demonstration is sufficient to generate comparable actions within the tolerance required by those coding this research (see Custance et al., 1995, for further discussion). When species are less-closely related, less careful body mapping is sometimes demonstrated (Custance et al., 1999). Even in human children, precise body mapping is only followed when the children assess it to be an important part of the demonstration (Gergely et al., 2002).

III. What makes humans different?

I now return to the question of why humans are unique in having exponentially accumulating culture. My explanation hinges on a difference in representational capacities.

Underlying Representations of Social Learning

The evidence of the previous section leads to the following conclusions about representation:

As discussed above, research into sense-action pairings as a basis for intelligent action is extensive in both AI and Cognitive Science. Forming and ordering these pairings may be a function of the hippocampus (Bryson and Leong, 2007).

The special case of vocal imitation in songbirds (and also parrots) has been the subject of extensive neuroscience research (see e.g. Leonardo, 2004). The upshot seems to be that a special neurological substrate is required, and it is not capable of learning and production at the same time. Since vocal imitation is not a common trait in nature, we must assume it evolves independently where it emerges (Marler, 1991), so bird results cannot necessarily be generalised to species like humans. I am not aware of similar neuroscience explanations of human vocal imitation. But because this representation may be a key to our cultural difference, I will review what information I have been able to find, which is largely due to Pöppel.

Pöppel (1994) documents a privileged representation of “phrases”, within which humans are capable of precise temporal memory and replication. These have a maximum duration of two to three seconds -- the exact duration seems to be under intelligent (though not deliberate) control and is situation-appropriate. That is, we tend to remember salient phrases of speech, music or gesture with appropriately-lengthed memory. The maximum possible duration of such episodes is presumably a cognitive constraint. Pöppel draws attention to the fact that most poetry and music consists of phrases of this length.

Implications for Cultural Evolution

That humans have this extra capacity while other primates do not is probably an accident of sexual selection (Vaneechoutte and Skoyles, 1998). As mentioned earlier, this accident may have provided us with a representation suitable for a memetic cultural-evolution explosion. Because so much more information is stored in the three seconds of detailed transcription than in the simple context-action pairs underlying programme-level imitation, knowledge represented in this domain can be highly redundant. This redundancy in turn can provide robustness where important data is stored, protecting it from an unsupervised process like evolution, and enabling operations analogous to cross-over and mutation to begin a full Darwinian process. I am discussing support and implications of this idea elsewhere (Bryson, 2008).

In the introduction of this article I mention another, currently unsupported hypothesis I am just beginning to research. This is that cultural learning is rare not because the mechanisms of learning required for an individual learner are difficult to evolve in themselves, but because of the impact on the ecological and social system supporting the learners. I believe that while cultural evolution has the potential to be a powerful means to search for new and more optimal behaviour; where cultural evolution exists, it must co-evolve with a set of constraints that damp its effects on the society and its ecosystem.

In the basic cases, this is obvious if you think about it -- if you are in a room with other people, look around yourself. Would it be a good idea if all of you converged on identical behaviour right now? The “crossover” equivalent in social learning, the mechanism of recombining “good tricks” (Dennett, 1995) from other conspecifics, must include a mechanism for maintaining diversity and supporting individual survival. One can also think about this in terms of longer-term consequences, such as population bubbles when a new, rich food source is discovered then driven to extinction.

Open Research Problems

Let us take the perspective that social learning is a risky strategy prone to positive feedback cycles that could result in a population's extinction, and that can only be stabilised with carefully co-evolved limits and damping mechanisms. Given this perspective, we can argue that we do see cumulative cultural evolution in a number of species (Franks and Richardson, 2006; van Schaik et al., 2003; Perry et al., 2003; Galef Jr. and Laland, 2005; de Waal and Johanowicz, 1993; Whiten et al., 1999). It just isn't accumulating as quickly as human culture, partially because the rate of change is actively damped by biological evolution.

This perspective leads to a number of open research questions, including:

IV. Summary

In this article I have proposed a very simple taxonomy of representational substrates for animal social learning: contexts, context-action pairings, and (for a very few species) short temporally-precise “scripts” of actions. I have also looked at the implications of this, both in relating social learning to well-known individual task-learning representations in AI and Cognitive Science, and for explaining why the sort of exponential growth in culture seen in human society is not witnessed so far in other species.

Joanna Bryson

Acknowledgements

Thanks to Andy Whiten and Mark Wood for many discussions of social learning and its representations.

References

Anderson, J. R., Matessa, M. P., and Lebiere, C. 1997. ACT-R: A theory of higher level cognition and its relation to visual attention. Human Computer Interaction, 12(4), 439-462.

Bishop, C. M. 2006. Pattern Recognition and Machine Learning. Springer, London.

Bryson, J. J. 2003. Action selection and individuation in agent based modelling. In Sallach, D. L. and Macal, C., editors, Proceedings of Agent 2003: Challenges in Social Simulation, pages 317-330, Argonne, IL. Argonne National Laboratory.

Bryson, J. J. 2008. Embodiment vs. memetics. Mind & Society. Accepted for publication.

Bryson, J. J. and Leong, J. C. S. 2007. Primate errors in transitive `inference': A two-tier learning model. Animal Cognition, 10(1):1-15.

Bryson, J. J. and Stein, L. A. 2001. Architectures and idioms: Making progress in agent design. In Castelfranchi, C. and Lespérance, Y., editors, The Seventh International Workshop on Agent Theories, Architectures, and Languages (ATAL2000) . Springer.

Bryson, J. J. and Wood, M. A. 2005. Learning discretely: Behaviour and organisation in social learning. In Demiris, Y., editor, Third International Symposium on Imitation in Animals and Artifacts, pages 30-37, Hatfield, UK. The Society for the Study of Artificial Intelligence and the Simulation of Behaviour.

Byrne, R. W. 1995. The Thinking Ape. Oxford University Press.

Byrne, R. W. and Russon, A. E. 1998. Learning by imitation: a hierarchical approach. Brain and Behavioral Sciences, 21(5):667-721.

Chater, N., Tenenbaum, J. B., and Yuille, A. 2006. Probabilistic models of cognition: Conceptual foundations. Trends in the Cognitive Sciences, 10(7):287-291.

Custance, D. M., Whiten, A., and Bard, K. A. 1995. Can young chimpanzees (pan troglodytes) imitate arbitrary actions? Hayes and Hayes (1952) revisited. Behaviour, 132:837-859.

Custance, D. M., Whiten, A., and Fredman, T. 1999. Social learning of an artificial fruit task in capuchin monkeys (cebus apella). Journal of Comparative Psychology, 113(1):13-23.

Davelaar, E. J. 2007. Sequential retrieval and inhibition of parallel (re)activated representations: A neurocomputational comparison of competitive queuing and resampling models. Adaptive Behavior, 15(1):51--71.

de Waal, F. B. M. and Johanowicz, D. L. 1993. Modification of reconciliation behavior through social experience: An experiment with two macaque species. Child Development, 64:897-908.

Dennett, D. C. 1995. Darwin's Dangerous Idea. Penguin.

Fitch, W. T. 2000. The evolution of speech: A comparative review. Trends in Cognitive Sciences, 4(7):258-267.

Franks, N. R. and Richardson, T. 2006. Teaching in tandem-running ants. Nature, 439(7073):153.

Galef Jr., B. G. and Laland, K. N. 2005. Social learning in animals: Empirical studies and theoretical models. BioScience, 55(6):489-499.

Gallistel, C., Brown, A. L., Carey, S., Gelman, R., and Keil, F. C. 1991. Lessons from animal learning for the study of cognitive development. In Carey, S. and Gelman, R., editors, The Epigenesis of Mind, pages 3-36. Lawrence Erlbaum, Hillsdale, NJ.

Garcia, J. and Koelling, R. A. 1966. The relation of cue to consequence in avoidance learning. Psychonomic Science, 4:123-124.

Gergely, G., Bekkering, H., and Király, I. 2002. Rational imitation in preverbal infants. Nature, 415:755.

Graziano, M. S. A., Taylor, C. S. R., Moore, T., and Cooke, D. F. 2002. The cortical control of movement revisited. Neuron, 36:349-362.

Henson, R. N. A. and Burgess, N. 1997. Representations of serial order. In Bullinaria, J. A., Glasspool, D. W., and Houghton, G., editors, Proceedings of the Fourth Neural Computation and Psychology Workshop: Connectionist Representations, London. Springer.

Hineline, P. N. and Rachlin, H. 1969. Escape and avoidance of shock by pigeons pecking a key. Journal of Experimental Analysis of Behavior, 12:533-538.

Hood, B., Carey, S., and Prasada, S. 2000. Predicting the outcomes of physical events: Two-year-olds fail to reveal knowledge of solidity and support. Child Development, 71(6):1540-1554.

Cerebral mechanisms in behavior. John Wiley & Sons, New York.

Leonardo, A. 2004. Experimental test of the birdsong error-correction model. Proceedings of the National Academy of Science, 101(48):16935-16940.

Marler, P. 1991. The instinct to learn. In Carey, S. and Gelman, R., editors, The Epigenesis of Mind, pages 37-66. Lawrence Erlbaum, Hillsdale, NJ.

Perrett, D. I., Mistlin, A. J., and Chitty, A. J. 1987. Visual neurones responsive to faces. Trends in Neuroscience, 9:358-364.

Perry, S., Baker, M., Fedigan, L., Gros-Louis, J., Jack, K., MacKinnon, K., Manson, J., Panger, M., Pyle, K., and Rose, L. 2003. Social conventions in wild white-faced capuchin monkeys: Evidence for traditions in a neotropical primate. Current Anthropology, 44(241-268).

Pöppel, E. 1994. Temporal mechanisms in perception. International Review of Neurobiology, 37:185-202.

Rizzolatti, G., Fogassi, L., and Gallese, V. 2000. Cortical mechanisms subserving object grasping and action recognition: A new view on the cortical motor functions. In Gazzaniga, M. S., editor, The New Cognitive Neurosciences, chapter 38, pages 538-552. MIT Press, Cambridge, MA, second edition.

Roper, T. J. 1983. Learning as a biological phenomenon. In Halliday, T. R. and Slater, P. J. B., editors, Genes, Development and Learning, volume 3 of Animal Behaviour, chapter 6, pages 178-212. Blackwell Scientific Publications, Oxford.

Santos, L. R. and Hauser, M. D. 2002. A non-human primate's understanding of solidity: Dissociations between seeing and acting. Developmental Science, 5. In press.

Seyfarth, R. M., Cheney, D. L., and Marler, P. 1980. Monkey responses to three different alarm calls: Evidence of predator classification and semantic communication. Science, 14:801-803.

Spelke, E. S., Breinlinger, K., Macomber, J., and Jacobson, K. 1992. Origins of knowledge. Psychological Review, 99:605-632.

Tanji, J. 1996. Involvement of motor areas in the medial frontal cortex of primates in temporal sequencing of multiple movements. In Caminiti, R., Hoffmann, K., Lacquaniti, F., and Altman, J., editors, Vision and Movement: Mechanisms in the Cerebral Cortex, volume 2, pages 126-133. Human Frontier Science Program, Strasbourg.

van Schaik, C. P., Ancrenaz, M., Borgen, G., Galdikas, B., Knott, C. D., Singleton, I., Suzuki, A., Utami, S. S., and Merrill, M. 2003. Orangutan cultures and the evolution of material culture. Science, 299(5603):102-105.

Vaneechoutte, M. and Skoyles, J. 1998. The memetic origin of language: modern humans as musical primates. Journal of Memetics -- Evolutionary Models of Information Transmission, 2(2).

Whiten, A. 1998. Imitation of the sequential structure of actions by chimpanzees ( pan troglodytes). Journal of Comparative Psychology, 112:270-281.

Whiten, A. 2006. The dissection of imitation and its `cognitive kin' in comparative and developmental psychology. In Williams, S. R. . J. H. G., editor, Imitation and the development of the social mind: Lessons from typical development and autism, pages 227-250. Guilford Press, New York.

Whiten, A., Flynn, E., Brown, K., and Lee, T. 2006. Imitation of hierarchical action structure by young children. Developmental Science, 9(6):574-582.

Whiten, A., Goodall, J., McGew, W. C., Nishida, T., Reynolds, V., Sugiyama, Y., Tutin, C. E. G., Wrangham, R. W., and Boesch, C. 1999. Cultures in chimpanzees. Nature, 399:682-685.

Wolpert, D. H. 1996. The lack of a priori distinctions between learning algorithms. Neural Computation, 8(7):1341-1390.

Wood, M. A. and Bryson, J. J. 2007. Skill acquisition through program-level imitation in a real-time domain. IEEE Transactions on Systems, Man and Cybernetics Part B--Cybernetics, 37(2):272-285.

Wood, M. A., Leong, J. C. S., and Bryson, J. J. 2004. ACT-R is almost a model of primate task learning: Experiments in modelling transitive inference. In The Annual Meeting of the Cognitive Science Society (CogSci 2004) , pages 1470-1475, Chicago. Lawrence Erlbaum Associates.

Zentall, T. R. 2001. Imitation in animals: Evidence, function, and mechanisms. Cybernetics & Systems, 32(1):53-96.