Modules for teaching AI with the Mechanical Neural Network

The teaching modules system allows the Mechanical Neural Network to be used in various teaching scenarios. Factors such as the age of the learners, their prior knowledge, the learning objectives, and the time frame are taken into account. Based on these factors, different teaching modules can be selected and covered in appropriate depth. Possible scenarios range from lectures at universities in the field of AI to short conversations at trade fairs or in museum contexts. In the first case, students have a lot of prior knowledge and relatively a lot of time, but may also need to understand and apply e.g. the backpropagation algorithm. In the second case, interested parties may have little or no previous contact with machine learning and AI and little time, but do not need to understand the mathematical details in depth. In between, there are various scenarios in schools where students can experiment with the Mechanical Neural Network . All of these scenarios are covered by the modular system. In the following the seven modules are outlined.

Module 1: Introduction

The first module introduces the basics of machine learning and AI. These include the concepts of supervised learning and unsupervised learning, classification and clustering, features, learning, training, test, and validation data sets, and data bias.

Building block 2: Assembly

In the second building block, the ANN is assembled or constructed. In the process, the learners develop an understanding of how the rockers/neurons move, how the rockers are connected, and how the clamps/weights can be moved.

Building block 3: Practical introduction

In the third building block, the basic concepts of artificial neural networks are introduced and explained using the ANN. These include neurons, network architecture, connections, weights, activation function, summation, bias, and forward pass. Learners also train the ANN: In a classification task, the ANN is supposed to learn to recognize animals based on various features. Learners understand what learning means for an artificial neural network and come to the realization that a planned approach is necessary for this.

Module 4: Excursus

The excursus can cover the history of and motivation for artificial neural networks. This includes logical perceptrons, nonlinear problems such as XOR, and the questions of why more than one neuron is needed and how many neurons are needed.

Building block 5: Backpropagation algorithm

In the fifth building block, learners understand and calculate the backpropagation algorithm. They enter their results into the MNN and check them. Several training steps are then run through until an example problem such as XOR is solved. In doing so, learners understand that training is an iterative process with local minima that depends on (hyper)parameters such as weight initialization.

Module 6: Transfer to the computer

In the sixth module, the fully connected network / multilayer perceptron is transferred to the computer. The computer-calculated results are validated on the MNN. Furthermore, deep learning frameworks are introduced and the MNN is recreated in the deep learning framework. After that, a more complex artificial neural network can be trained on a sample dataset such as MNIST in the deep learning framework.

Module 7: Classification

In the seventh module, the MNN is embedded in the context of the latest AI developments. Possible topics include task-specific neural networks, large language models, foundation models, general-purpose AI, and AI safety.