AS Logo

The Mechanical Neural Network(MNN)

Adjusting the weights of the MNN to adapt it to the XOR problem

The Mechanical Neural Network is a mechanical implementation of an artifical neural network. To be more specific a Multilayer Perceptron with ReLU activation functions. It can be adapted to model real valued function or logical function, like the logical AND, OR, and the exclusive or(XOR). This means that we can adapt the weights between the neurons of the network such that for a given output the network produces the correct output. E.g. we can adjust the weight such that the network models the XOR functions and outputs true if and only if one of the two inputs is true and the other false.

It is build from wooden levers, which represent the neurons. This levers are linked by strings, corresponding to the connections between the neurons of a MLP. Adjusting the clamps, which connect the strings to the levers, by hand allows to adjust the weights of the network. Thus, the effect of adapting the weights can be intuitively observed.

For more information you can watch the Introduction Video on YouTube:

Or read one one of these papers:

Promoting the digital transformation of STEM education with the mechanical neural network, a physical model for future-oriented and student-centered AI education
A. Schaffland, C. Müller, and J. Schöning, 2024, IEEE Global Engineering Education Conference (EDUCON), IEEE, DOI: 10.1109/EDUCON60312.2024.10578714
Mechanical Neural Network: Making AI comprehensible for everyone
A. Schaffland and J. Schöning, 2023, IEEE German Education Conference (GeCon), IEEE, DOI: 10.1109/gecon58119.2023.10295144

News about the MNN are presented here.

And here are some more images: