Project DetailsArtificial neural networks are advanced computational systems designed to simulate the functioning of the human brain, yet their conceptual complexity often makes them difficult to grasp intuitively. The goal of this project was to translate the theoretical idea of a neural network into a tangible physical model based on water flow, demonstrating how information propagates between neurons within the network.
The developed system illustrates the principle of forward propagation - the process through which data passes from the input layer to the output layer via hidden layers. This model provides a physical visualization of how a neural network performs classification tasks, in this case recognizing handwritten digits from the MNIST dataset, by representing differences in “water” volume flowing through the channels.
Instead of abstract numerical representations of weights and activations, the water-based network expresses computation through fluid movement, turning a mathematical process into a clear visual experience. The model enables direct observation of how each layer operates, how internal connections influence the flow, and how adjustments to the weights affect the final classification output.
The project combines physical principles of fluid dynamics with data processing and machine learning, offering a novel interpretation of neural networks, not only as computational tools but also as educational models that make the core concepts of artificial intelligence more accessible and intuitive. By merging analytical precision with physical embodiment, this approach enhances understanding of how computational processes can manifest in the real world.
