A colony of single-celled organisms can function as a biological computer to crunch a series of historical data points and forecast the future. An experiment that replaced each node in a neural network with a tiny organism showed that the waxing and waning size of the colony could accurately forecast the next step in time series data, such as the size of daily fishing catches of certain species in Japanese waters.
Neural networks are a highly popular form of artificial intelligence that uses clusters of mathematically connected nodes to mimic the way that neurons and synapses function in a real brain. The strength of connections between nodes is tweaked over hundreds or thousands of training runs in order to teach the network how to perform certain tasks or spot certain patterns, just as the synapses of the brain reinforce certain patterns when we learn a new skill or memorise new information.
is a special type of artificial neural network where the strength of connections between nodes remains fixed, so that a given data input will always cause the same chain reactions through the artificial brain. It is the readout from the reservoir computer, rather than the connections between the nodes in the network, that the AI learns to tweak. This is then used to predict the next step in the data.
Advertisement
A big advantage of reservoir computing is that it doesn鈥檛 require the network of nodes to constantly change during learning, which makes the networks easier to build and train. As such, researchers have spent the past few years exploring using various physical materials as a reservoir computer. Research has indicated that the use of photons, analogue electrical circuits, fluids and biological material as a reservoir that can perform calculations is promising.
at Kyoto University in Japan and colleagues have now shown that colonies of a single-celled eukaryotic organism called Tetrahymena thermophila, each just 0.1 millimetres long, also function as a reservoir and can be used to make accurate forecasts.
The team created an incubation chamber with a clear bottom and introduced a colony of T. thermophila. The chamber was placed under a microscope and a camera took an image of it once a minute, which an algorithm used to automatically count the population.
In the experiment, the chamber and organisms formed the reservoir computer. The input data 鈥� in this case, information on the size of catches from fishing vessels 鈥� was converted to a series of temperature readings. The incubation chamber鈥檚 temperature was adjusted accordingly, then changed after a set time interval to the next temperature in the series. Data on the size of the growing and shrinking colony was the output. Although the number of organisms grows and shrinks, no changes are made to how they interact and the reservoir is left to its own devices as a self-contained unit.
Multiple experiments revealed that the colonies reacted predictably and similarly if given the same set series of temperature inputs, despite starting from non-identical states. This is a defining feature of a deterministic system and suggests that the set-up had potential as a computer.
鈥淚 expected that they should synchronise, at least to some extent, but this level of synchronisation was beyond my expectation,鈥� says Ushio. 鈥淭his means that although it鈥檚 a very complex and very complicated system, they are still deterministic. This is a very important property necessary for computation.鈥�
The reservoir used the available data to predict between 15 and 30 data steps ahead, using its predictions as inputs for each step into the future. In two experiments, the computer was asked to forecast catch estimates for fishing vessels targeting flatfish or Japanese jack mackerel. The predictions made by the T. thermophila were more accurate than those made by mathematical forecasting models at certain time points when compared with real data, says Ushio, although the system is currently much less accurate and powerful overall than a traditional software-based reservoir computer.
鈥淭o me, it is not surprising that a biological system can perform computation,鈥� says Daniel Gauthier at the Ohio State University 鈥� but he says it isn鈥檛 yet obvious how such a system would prove useful in the real world.
Although the experiments aren鈥檛 yet peer reviewed and the T. thermophila show a greater ability to predict at certain timescales than others, Ushio believes that there is scope for improvement with different types of biological reservoir computers.
Reference: bioRxiv:
听
