Research

"Look on my works, ye Mighty, and despair!"
- Shelley's Ozymandias

"What a waste of scientist talent, that's if these jokers have any."
- jimonline, comment on Mail Online

Fish avoid predators by moving more randomly. If a prey fish detects a predator at a distance it has a good chance to outrun it by swimming directly away. At a closer distance the predator may be able to catch the prey if it swims in a straight line, so it makes sense for the prey fish to use turning and speed variations to escape the predator. In particular, the prey needs to move in ways that are unpredictable to the predator in order to avoid interception. In our paper in the Journal of Experimental Biology (Herbert-Read et al. 2017) [pdf] we show that this prediction is borne out in experiments. By measuring the complexity or 'randomness' of the escape path of a fish exposed to a artificial predator, we show that when the predator is closer the prey fish moves more erratically, potentially confusing the predator and giving it a chance to escape.

Optimising collective intelligence with incentives that increase diversity. What do democracy, stock markets, juries and academia have in common? Each depends on the collective wisdom of a group, community or society to make better decisions than a single individual. They can do this because each person brings different information, experiences and perspectives, resulting in a collective intelligence that surpasses any individual alone. But crowds are not always wise. In recent years, prediction markets have failed to predict either Brexit in the UK or President Trump in the USA. In this paper published in PNAS (Mann & Helbing 2017) we show that the way individuals are rewarded for the accuracy of opinions they express has powerful effects on group intelligence. Traditional market rewards (like in a stock or prediction market) tend to encourage everyone to look at the same few data sources. We show that a new reward scheme, designed to reward accurate minority opinions, restores collective wisdom to optimal levels.
Read more: The Conversation, Nature Physics

The importance of maintaining a diversity of approaches in data science. As with many academic fields, data science is subject to waves of fashionable and unfashionable research. In recent years, for example, Deep Learning and complex neural network models have become extremely prominent across most venues publishing research in machine learning. In this article, published in Data Science (Mann & Woolley-Meza 2017), we give our viewpoint on why it is important to maintain a diversity of modelling approaches as an academic community. Using a combination of network science and evolutionary game theory we suggest how we can adapt the structures and incentives that researchers work within to achieve the goal of more diversity and better collective wisdom.

Picking the best collections collectively without communication. Imagine you and your friends go to the shops to buy ingredients for a meal. You need to pick a set of ingredients that will work well together, but you lose each other in the shop and cannot contact each other to coordinate what you should each buy. How can you decide what to get so that the group's meal will be as good as possible? In our paper presented at the Conference on Autonomous Agents and Multiagent Systems (Malkomes et al. 2017) we address problems like this. We investigate algorithms for optimising group collective decisions with communication between agents and apply these algorithms to problems of computational search tasks, automated environmental monitoring and social insect foraging.

Fish make more use of social information when predation is low. Several rivers in Trinidad have a special property that allows the effects of predation to be studied: the lower reaches of the river, nearer the sea, have been colonised by predatory species that originated in the ocean. Higher reaches are protected by the existence of waterfalls that have stopped these species from migrating further upriver. As a result there are spatially-separated populations of guppies (a small prey fish) that have over a long time been exposed to either high or low predation risk. We studied how these different populations used social information when foraging. We found that while guppies from low-predation populations are able to improve their individual decisions about where to forage when in a group, those in high-predation populations were not (or did so sufficiently little as to make any effect statistically non-significant). Our results, published in Behavioral Ecology (Clément et al. 2017) suggest that there is a cognitive trade-off between predator detection and social information use.

Focusing on small data sets to assist hospitals with bed management. There is a lot of excitement about the potential for Big Data to revolutionise the delivery of healthcare, but often a wide gap between the hype and the implementation. In this paper, published in Frontiers in Public Health (Mann et al. 2016) we look at how the unglamorous but essential task of bed management can be assisted by using the data routinely collected at hospitals to predict how long patients will need to stay after an operation. Making use of this existing data could prevent cancellations of elective operations and reduce the number of beds needed to guarantee emergency treatment, saving patients and the NHS time and money.

Problem solving without a brain. All organisms have to solve problems such as how to find food and avoid predators. Some problems can be solved using built-in rules of behaviour, like the way plants can grow towards sunlight. Other problems require learning from past experiences, like remembering your route home from school or work. We learn through changes in the connections in our brains, but what about organisms that have no brain? Can they learn? In this paper in J R Soc Interface (Reid et al. 2016) we show that slime moulds, a simple single-celled organism with no neurons, can solve a classic learning task, the Two-Armed Bandit game. The cell expands more in areas where it has previously encountered food, resulting in an eventual decision to grow only towards higher food sources that mimics simple learning rules for playing the game.

Read More: TreeHugger, Phys.Org, IFLScience
Watch: Simon Garnier describes his lab's slime mould research

Read More: Heapcraft Project website.
Watch: Video demonstration

Visualising and analysing player behaviour in Minecraft. In Minecraft players inhabit a virtual world where they can explore, socialise, fight or collaborate to create new structures and experiences. Using our Heapcraft framework for data analysis and visualisation, we provide tools to help players and server administrators create the worlds they want, and effectively manage their virtual societies. In our paper presented at Foundations of Digital Games (Müller et al 2015) we demonstrated these tools and point the way to using data from Minecraft servers to answer social science questions about how societies are formed and maintained, and what types of social structure enable effective interaction and co-operation. Our upcoming paper at Motion in Games gives full details of all the publicly available tools our project has made available, while our paper at Artificial Intelligence and Interactive Digital Entertainment we show how collaboration between players can be measured and improved using our tools.

Causal entropic forces create collective organisation. Following the discovery that systems that maximise their entropy or 'variability' over possible future paths in an uncertain world exhibit intelligent behaviour, we assessed what this type of intelligence implies for collective self-organisation. In our paper in J R Soc Interface (Mann & Garnett 2015) we show that groups of agents who maximise their long-term entropy over possible movements or decisions in the future will naturally interact with each other, using social interaction rules such as Weber's law and 'social forces' that closely match many observed interactions in real animals and humans. This suggests that such groups can be understood via aggregate entropic methods without considering the detailed interactions between group members.

Large groups use negative feedback to avoid overcommitting to one option. Although they are single-celled organisms, slime moulds often act like groups when making decisions. Each part of the slime mould chooses to expand or move based on local cues, creating decision-making behaviour that closely resembles a school of fish or a flock of birds. Collective decisions are often driven by positive feedback - each individual becomes more likely to choose an option if others have already done so. This can often lead to information cascades where the whole group commits to one option, which can lead to other choices being under-exploited. In our paper in PLoS Computational Biology (Zabzina et al. 2014) we show that slime moulds are able to avoid this outcome when they have a large mass (equivalent to a large group), by using negative feedback to prevent too much mass being committed to any single option. Technically this negative feedback creates a new bifurcation for large masses/groups that introduces a new stable state where the slime mould divides itself between options. These results suggest that such mechanisms may regulate group sizes by favouring the splitting of very large groups, and prevent too many individuals trying to exploit the same resources.

Interaction rules, preferred positions, and noise. Previous studies have used a data driven approach to identify interaction rules between individual animals - how each individual changes their movement in response to others. These studies looked for attraction rules - moving towards the neighbour - and alignment rules - moving in the same direction as the neighbour. Our paper in Movement Ecology (Perna et al 2014) shows that attraction towards a preferred position relative to the neighbour can look like an attraction rule or an alignment rule, depending on where the preferred position is. We also show that "noise", either from tracking errors or the animals' intrinsic variable motion can create the illusion of particular interactions even where none exist. Together this means that future studies need to be more careful in how interaction rules are defined and identified.

Inside the mind of a sheepdog. The use of dogs to herd animals is a millenia-old tradition, which is still active today on the farm, the hills and in competitive sheepdog trials - even on television through shows such as One Man and His Dog. That such an established farming practice can make it to our TV screens speaks to the impressive spectacle of watching a solitary dog herding dozens, or even hundreds of sheep, seemingly moving the whole herd at will. How can one animal exert such precise control over so many others? What rules does it use to make sure it moves the herd in the correct direction, and never loses individual sheep? In our paper in J R Soc Interface (Strömbom et al. 2014) we investigate the rules the sheepdog uses, using a combination of agent-based models and data recorded from tracking sheep and dog in real herdings in Australia. Our results show that just two basic rules, driving and collecting, can explain the success of the sheepdog, and in the process naturally explains behaviours such as zigzag motion that previously were assumed to be explicitly learnt by the dog. Photo © AJ Morton

Read more: BBC News, The Times, Independent, Guardian, Daily Telegraph, Daily Mail, Sydney Morning Herald, Speigel Online
Watch: Video comparison of data and model

Foraging birds choose sites popular with both their own species and other species. In our paper published in Animal Behaviour (Farine et al. 2014) we monitored individual birds trips to feeding stations using radio tags. The birds were more likely to feeders that already had several other birds feeding on them. By fitting a Bayesian decision-making model we assessed how likely individual birds were to pick one of several feeding sites, based on the number of other birds at the site, both conspecifics (birds of the same species) and heterospecifics (other species). While the attraction to an individual of the same species was stronger, the larger number of heterospecifics means that in aggregate the birds of influenced just as much by the presence of other species. Our results show that a simple attraction rule can be a flexible way of finding good food sources in the wild, by exploiting the information provided by other foragers.

The complexity of university education. The performance and experience of university students depends on many factors, which themselves are strongly dependent on each other. Examples include the amount of time devoted to study and the need to work to finance their studies, which both depend on where the student can live and the price of accommodation. In this paper in PLoS ONE (Forsman et al. 2014) we use complexity theory and network analysis to analyse how all these factors influence each other, and how interventions to change any one of the many aspects of student life are likely to change the average performance of students in their examinations.

Mixed personalities are best for foraging groups. The 'personality' of an animal group emerges from the individuals within the group and their interactions. The mixture of individual personalities within the group may lead to overly conservative groups where no-one is willing to take the lead, fractious groups where no-one can agree, or some form of happy medium. In our paper in Proc. Roy. Soc. B (Aplin et al. 2014) we identify different personalities in groups of wild birds, find that birds with different personalities have different social interactions, and then show that a mixture of different personalities is essential to optimise the groups foraging behaviour - no single personality type can effectively explore and exploit the available food sources. These results suggest that we should look to maintain diversity of personality in both animal and human groups.

The dynamics of human development. What causes societies to become more democratic, to increase their people's life expectancy, education and well-being? Is greater education necessary for the development of democracy, or is "people power" required before the government will provide education for the masses? How is a society's administration related to the individual values of its members? Are democracy and permissive values a luxury that only wealthy countries can afford? By applying our previously developed methods to the study of data from the World Bank and the World Values Survey we uncover the dynamical relationships that control changes in individual and societal values and development in this paper published in PLoS ONE (Spaiser et al. 2014)

Watch: Model simulation predictions for India and Ukraine

Hedges and edges help pigeons learn their way around. How do pigeons memorise their way home? By analysing a combination of actual flight paths from homing pigeons, recorded with GPS trackers, and images of the landscape around Oxford where these birds were flying, we try to answer this question in our paper in Biology Letters (Mann et al. 2014). We show that the navigating pigeons were able to return more accurately to areas of the landscape with a special property - these areas contain an optimal amount of 'visual information'. Areas with very little structure, such as open fields, proved difficult to memorise. Likewise very complex landscapes such suburban Oxford were also a problem. But in between lies a 'sweet spot' where the pigeons were best able to recognise and return to locations they had seen before. These areas (shown left), often lying between town and country, or between field and forest, are likely to be key in understanding the foundations of birds visual memories.

Watch: SVT Vetenskapsvärlden, ITV News
Read more: Discover Magazine, Daily Mail, Daily Telegraph, Irish Independent, SVT, Navigation News

Bringing clarity to international development. Our study published in PLoS ONE (Ranganathan et al. 2014) shows how raw data about countries' development, collected by the World Bank and other organisations, can be analysed to discover the relationships that drive economic and social progress. Applying Bayesian methods to automatically select between hundreds of competing models we find the set of differential equations that best explain the development seen in GDP, democracy, child mortality and other measures over the past decades in countries from across the globe. Understanding whether economic growth drives democracy or vice versa, whether education or health is more important for further economic development, and whether there are crucial tipping points in a country's development means we can better predict and assist the path of development in the future. Check out the accompanying R package to apply these methods yourself

Watch: Video abstract Read more: IFFS press release, The Development Space project

Fish follow the 'most recent mover' when making decisions. The positions and movements of other individuals provide fish with a valuable source of social information. In particular, the current locations of other fish can indicate where food might be found, and where it is safe to stay. We expect therefore to find fish moving towards larger groups, to take advantage of the safety and information they provide. However, in our experiments, published in J R Soc Interface (Mann et al. 2013), we find that humbug damselfish used different strategy when choosing to move between two coral areas. Instead of trying to find the larger group, each individual simply copied the actions of the last fish it saw moving, leading to cascades of fish moving between the two coral. Such a strategy may help the fish know when it is safe to explore, or to detect when it is best to try new options.

Read more: David Sumpter's Blog

Crowds are wise for hard tasks, but not for easy ones. 'Swarm intelligence' and the 'wisdom of the crowd' are commonly used terms suggesting that groups of individuals can outperform individuals in a variety of tasks. Conversely, the 'madness of crowds' is often invoked to explain disastrous group behaviours, like fatal stampedes and stock market crashes. Our results, published in PNAS (Sasaki et al. 2013), show that ant colonies are much better at solving hard problems than individuals on their own. However, when tasks are relatively simple (such as choosing between a great nest site and a terrible one), individual ants perform better on average than the colony, as the group sometimes blindly follow each other rather than assessing the nest quality themselves. Photo © Takao Sasaki

Read more: Nature, National Geographic, The Atlantic, LA Times

Clapping spreads as a social contagion. Our study in J R Soc Interface (Mann et al. 2013) uses Bayesian model selection to show how applause spreads through an audience like a disease, with each individual being 'infected' by the social pressure created by those already clapping. The end of the applause is similarly mediated, with one or two individuals stopping through boredom or tiredness, followed by a wave of other audience members who respond to the increasing pressure to stop. The best model for how clapping starts, spreads and stops suggests that performances of identical quality can generate very different durations of applause.

Read more: Science Magazine, BBC News, The Times,The Guardian, Daily Mail, Slate, Wissenschaft, SvD
Listen: BBC WS Why Factor, BBC R4 Seriously, BBC R4 Material World, NPR, Scientific American, Radio New Zealand

Memory of interactions can drive collective motion. Collective motion, such as the swarming patterns of flocks of starlings, or the rotating ball of a school of sardines, is generated by the individual movement choices of the individuals in the group. These choices are driven by the interactions of each individual with the animals around it. Much debate has centred on the question of which neighbouring animals each individual pays attention to - those within a fixed spatial range, a certain number of nearest-neighbours, or only other animals the individual can directly see? In this study in PLoS Comput. Biol. (Mann et al. 2013) we show that in addition to these spatial limits on interactions, the temporal range must also be considered. Observing the collective motion of glass prawns, we use Bayesian model selection to show that the individual's memory of its recent interactions is important in determining its future direction choices.

Male fish create food-shaped lures to attract females. Male sword-tail charagins have protruding appendages that they use to attract the attention of females, luring them into a position from which they can transfer sperm. It had been suggested that these lures may be attractive to the females because they resemble possible food items, much like a fisherman's 'fly'. In this study published in Current Biology (Kolm et al. 2012) we showed that female fish were more attracted to lures resembling their habitual food sources. After only 10 days on an ant-based diet, females were more attracted to lures that resembled ants, demonstrating both the accuracy of the food-lure hypothesis, and the rapid adaptability of the females' preferences.

Read more: Science Magazine, New York Times, Scientific American, Daily Mail,Huffington Post, Discover Magazine
Listen: Sveriges Radio

How we model collective animal behaviour. In this review, published in Interface Focus (Sumpter et al. 2012) we describe the process of modelling the behaviour of groups of animals, focusing on the interplay between what happens on the individual level and the behaviour of the group as a whole. Based on recent work in various areas of collective animal behaviour, we give examples of how the individual and group level behaviour have been approached, and suggest a cycle of theory-driven and data-driven research to find models that accurately describe how the two are connected.

An optimal strategy for playing Battleships. In this paper, presented at ICML (Garnett et al. 2012), we develop a theory for how to perform optimal binary search. Among many real-world applications for this method, including fraud detection, terrorist identification and resource location, one simple example is the game of Battleships. In this game the players aim to sink each others ships, which are hidden from view, by placing 'shots' into nominated grid spaces. As the players hit or miss with their shots they learn about the possible locations the other player has placed their ships in. By designing a strategy for a sequence of shots which maximise the amount of information the player receives, and thus increases their chances of making as many successful shots as possible, our method can be used to play the game in the most effective way.

Watch: Roman Garnett presents the paper here

Fish adapt their speed and heading in response to their nearest-neighbour. We recorded the movements of shoals of mosquitofish, using video tracking to reconstruct the trajectory of each fish in the group. Using a neural networked based learning algorithm we identified the factors that were responsible for forcing each fish to adjust its speed and direction. Our results, published in PNAS (Herbert-Read et al. 2011) showed that fish slowed down or sped up to stay in contact with the group while avoiding collisions, and turned towards their closest neighbour to maintain group cohesion. Both the speed adjustment and the direction changes were primarily the result of interactions with the closest individual to the focal fish.

Watch: Video of one of our experiments with 4 fish shoaling
Read more: Daily Telegraph , ABC , Discovery, Australian Geographic, My blog posts on this topic

How to test models of collective motion. There are a lot of different models that are used to describe how animals move together in groups. Because the behaviour of the whole group is an emergent property, different rules describing the interactions between individuals can often produce very similar group movement. This makes it difficult to test how the individual animals are interacting with each other. In this theoretical paper, published in PLoS ONE (Mann 2011), I showed how we can use recorded movements by individual animals to find out which of the current models are the best description of these individual interactions, and what some of the challenges we might face are as we put this into practice.

Information theory reveals how pigeons learn their way home. By repeatedly releasing homing pigeons from the same location and using GPS tracking technology to follow their flights home, it could be seen that these birds memorised a route from the release site to the home loft, which they followed on each flight with increasing accuracy. By using Gaussian processes and information theory we were able to quantify these route memories in terms of the amount of information they contained, and thus establish how quickly the birds learnt their routes in this study published in J R Soc Interface (Mann et al. 2010). By isolating the parts of the route that contained the most information we could also identify which parts of the landscape were being used as landmarks. Photo © Robin Freeman

As part of my PhD thesis I extended these results to show which birds were sharing or receiving information about the route home when flying in pairs.

Read more: My blog posts on this topic, Navigation News

Pigeons choose better navigators as their leaders. Studying the flights of pigeons released together in pairs shows that some birds consistently tend to follow others home. By analysing how well each individual had memorised its own way home before being paired with other birds, we found that the birds with the best memories of their own preferred routes tended to be the leaders, while those who were more uncertain about their own routes tended to follow their partner. The results were published in Biology Letters (Freeman et al. 2010)
Read more: The Conversation

Changes in behaviour of foraging macaroni penguins. Foraging animals typically alternate between periods of active feeding and periods of exploration and travel. Understanding how the animal's conditioning, recent history and life-phase affect the foraging strategy has been restricted by difficulties in accurately identifying the transitions between the different behaviours. In this paper in Marine Biology (Hart et al. 2010) we were able to overcome some of the statistical difficulties inherent in this task, using a Hidden Markov Model approach. Applying this technique to depth recordings of foraging macaroni penguins, we were able to relate the amount of time spent feeding versus exploring to the reproductive cycle stage of the individual penguins

Photo © Jerzy Strzelecki (CC BY 3.0)