History
The development of the Mechanical Neural Network is based on two questions:
- How can we explain AI and artificial neural networks to people without computer and programming skills?
- How can analog artificial neural networks function?
The answer to the first question is not only interesting for people
without computer and programming skills, but also for those with a
background in computer science. Initially, the aim when teaching with
the Mechanical Neural Network was to separate understanding,
calculation, theory, and programming from one another, and to develop a
different perspective and approach to understanding. The students were
thus able to experimentally apply the theory learned in the lectures in
practice with the Mechanical Neural Network and set and check their
calculations of forward pass and backward pass with the backpropagation
algorithm on the model. This, in turn, led to the development of
concepts for workshops and the modular teaching system, taking into
account the experiential learning model. 
For the second question, there are various possible solutions, including
analog electronic circuits, digital circuits, and purely mechanical
approaches. The mechanical solutions are the most intuitive ones. The
flow of information and the training of the network are clearly
visualized through the movement of the individual components. No
knowledge of electricity or electronics is required.
The first Mechanical Neural
Network was made entirely of wood, with the seesaws made from CNC-milled
parts that were glued together. The entire frame and base were also made
of wood, and assembled using handcrafted wood joints. In the following
iterations, the design switched to 3D-printed parts, which can be
manufactured more precisely and more quickly. As part of this change,
the input neurons were equipped with a ratcheting mechanism, and the
clips that model the weights were replaced by sliders that also lock
into place on the seesaws. This allows both the weights and the neurons
to be adjusted more easily. Another improvement is the use of
multicolored strings, which make the movements, and thus the flow of
information through the network, easier to visualize.
The modular system is currently being further developed, supplemented with worksheets and videos, and continues to be evaluated.