A good test of how well we understand something is how well we can reproduce it on a computer.
This type of 'computer simulation' test was proposed by mathematician Alan Turing as a way of determining whether scientists had created Artificial Intelligence. He suggested that if we can successfully simulate human behaviour, so that a person interacting with a computer thinks it is interacting with a human, then we have passed the test. The philosophy behind this idea is that when two things behave the same, then to all intense and purposes, they are the same.
Although there have been some recent high profile claims that the Turing Test has been passed, most scientists think that it will be quite a long time until we build a computer program that fools us in to thinking it is human. All you have to do is talk to iPhone’s Siri for a minute or two to realise how far we are away from solving the problem. While computers have developed rapidly since Turing’s time, progress on General Artificial Intelligence has been slow.
If we can’t simulate humans, can we simulate other animals? Over the last five years our research group has been trying to understand how fish and other animals move in groups. Part of this process is building simulation models, where we program 'particles' to move with the same rules of motion as we observe in the fish. The aim is to see if we can understand how fish interact to produce schooling patterns.
Usually, we test our models by statistical comparison to data. On the face of it, this might appear like the most 'scientific' way of checking whether the model works. But it turns out to have an unexpected limitation. Often models will pass a series of statistical tests, but when we look at them on our computer screens they don’t really look like real fish. A human comparing the movement of simulated fish particles and real fish can immediately see the difference.
Recently, we created a model of Pacific Blue Eye fish. We have already subjected the model to a series of statistical comparisons to the data, and it works well. But can the model fool a human in to thinking it is real fish? Can it pass a fishy version of the Turing Test?
To find out, we first filmed fish in arena and used a computer to track the position of all the fish. We then created a simulated sequence of fish movements. By displaying dots to show the positions of the real fish and the simulated fish at the same time we built a spot the difference test. You can try the test yourself, by clicking on the link below. You will see two annimations at the same time, one is real fish, the other is simulated. Can you tell which is real?
We think it will probably be pretty difficult for those that aren’t used to watching fish in the lab to tell the difference, but we’ll see. We are currently collecting data on how people visiting this site perform in the test.
The real test of our simulation is whether it can fool fish behaviour experts. At a recent workshop in Uppsala we asked the visiting researchers, all of whom work on experiments and models of fish schools, to do the test. They are a competitive bunch and were highly motivated to show that they could tell the difference between real and simulated fish. And they had lots of fun trying to spot limitations in our simulations.
Below are their results. They each played the game 4 or 5 times, each time looking at 6 clips of real and simulated fish. A few of them got 6 out of 6 first attempt. On average they were slightly better than random, getting 4.5 out of 6. Surprisingly, they got slightly worse on the second and third attempt to spot the difference. But by the 4th and 5th attempts it looked like they cracked it. Most were getting 5 or 6 out of 6. It is reassuring that the fish biologists aren’t fooled by a computer program.
The results do however mean that we have still got some work to do if our fish simulations are going to pass something like a Turing Test, on experts at least. We are curious how our website guests perform. In a few weeks we’ll publish the results and tell you how you performed as a group. We can’t and don't record individual identities: we wouldn’t want to embarrass any fish experts that play. So get playing.
Model and Simulation: Maksym Romenskyy
Data collection: James Herbert-Read
Text: Maksym Romenskyy & David Sumpter
Turing, A. M. (1950) Computing Machinery and Intelligence. Mind 49: 433-460.
Sumpter, David J. T., Richard P. Mann, and Andrea Perna. The modelling cycle for collective animal behaviour. Interface Focus (2012)