I made a neural network that I can train using my fingers (medium.com)

A tech enthusiast has created a unique physical neural network known as KnobNet, which allows users to train the model using physical knobs instead of conventional digital methods. Unlike traditional neural networks that utilize backpropagation to adjust weights and biases automatically, KnobNet requires manual input, where the user turns knobs to modify settings based on immediate feedback from a loss screen. This hands-on approach deepens the understanding of how neural networks learn, providing an experiential learning platform where the creator acts as the learning algorithm. The significance of KnobNet lies in its ability to demystify the training process of neural networks by making it tangible and interactive. Each adjustment can be observed in real-time, revealing the interconnected impact on recognizing digits from 0 to 9. The project involved several technical aspects, including the design of printed circuit boards, use of a Raspberry Pi 4B for operation, and crafting a user-friendly GUI. While this manual method is slower than traditional computing techniques, it serves as an educational tool that offers insights into machine learning fundamentals. The success of KnobNet highlights the potential for innovative teaching methodologies in AI/ML education.