Because it is the most efficient method that we have available to us, artificial neural networks are almost always instantiated as mathematical models on digital computers. But in the same way that a computer need not be composed of transistors, but could just as easily be made from vacuum tubes, relays, or gears, neural networks can also be created in alternative ways. They can even be built into physical structures that are engineered to interact with one another, much like the layers of a mathematical model, such that they can transform an input into a prediction.
There may not be very many practical reasons to do such a thing, but Dietmar Offenhuber and Orkan Telhan have found that there is at least one. Called the Reservoirs of Venice , their art installation was designed to educate, entertain, and spur creative thinking. It consists of columns of water that interact with one another. Through these interactions, the system has learned to predict the time of day based on human activity that is observed near the canals of Venice.
As an input, the network takes in information from webcams that are positioned around the canals of Venice. Information about human activity is extracted and translated into water disturbances, which are created by wave makers in a set of four water reservoir columns. The disturbances in one column influence the disturbances made in the next, and that process repeats until the final column is reached.
The water disturbance from the last column is then fed into a traditional single-layer digital neural network. It is this segment of the network that interprets the data produced by the physical components of Reservoirs of Venice to determine what time of day it is, which it learns to do after seeing many examples.
The water reservoir columns could quite obviously be replaced by a few more layers in a digital model. This would make things a whole lot simpler to build, and would also make the computations a great deal faster. But with the water columns, it makes it a lot easier to visualize what is actually happening inside the neural network. It is also a good reminder that the way we do things now is only one option, and perhaps not even the best. And different thinking is exactly what we will need to build smarter and more energy-efficient solutions in the future.
The Reservoirs of Venice is a physical neural network (📷: D. Offenhuber and O. Telhan)
The input layer (📷: D. Offenhuber and O. Telhan)
The output layer predicts the time of day (📷: D. Offenhuber and O. Telhan)