Reservoirs of Venice is the prototype of an urban-scale computer that works as a physical neural network. It uses water waves to learn from patterns of human activity to understand long-term changes in the environment.

Unlike biological neural networks found in most animals, or digital neural networks used in AI, this system uses the material dynamics and memory of water waves. The project invites us to see the Venetian canals as a physical neural network that processes environmental information. It proposes a new model of ecological urban computing that leverages the dynamics of urban and environmental processes.

The physical neural network installation consists of six columns that serve different functions. The screens column collects information about human activity on the bridges and canals of Venice from a network of webcams around the city. This information is passed into the four water reservoir columns, which act as a neural network. Each column uses water disturbances to transform the input data and feed it to the next column. Finally, the time column uses a single-layer digital neural network to interpret the recursively transformed data and learns to infer the corresponding time of day.

Reservoir computing is a new area of computer science that aims to reduce the energy required to train deep neural networks by keeping most layers static, forming a "reservoir" of digital neurons, and training only the final output layer that interprets its transformations. Physical reservoir computing replaces this reservoir with physical processes to achieve equivalent nonlinear transformations. The temporal dynamics of the physical medium acts as a short-term memory for the computation: while the waves are chaotic and unpredictable, the time it takes for them to build up and settle is invariant.

Over time, the installation learns what a typical morning, afternoon, or evening in Venice looks like: the flow of pedestrians and dogs, the vaporettos, water taxis, and boats. In this sense, the installation is a model for an urban reservoir computer: the water surfaces of the canals can be compared to how information moves in the neural network of a reservoir. If we learn to process its patterns, we can use the activity of the city to reflect and interpret its own state.

There is a long history of water computers, from antiquity to Al-Jazari's Water Clocks in the 13th century, Vladimir Lukyanov's Water Integrator of the 1930s, or Bill Phillips' MONIAC of the 1950s, a machine that simulated the national economy. They all use water as an analogy for the quantities to be calculated: time, temperature, or money. Our physical neural network does not need analogies, but uses the expressions and memory of water as an elemental medium.