Show HN: Distributed Training via Webcams (www.sarthakmangla.com)

At the Neo Hackathon, a team comprising Simon, Pav, and Faisal developed a unique approach to distributed training using webcams as a communication medium, a stark deviation from traditional methods. They set up laptops where each webcam could see every screen, transmitting data visually without WiFi or wires. The experiment aimed to push creative boundaries within a limited timeframe, achieving a functional, albeit impractical, mechanism for machine learning model training using the MNIST dataset. They designed a system where training steps were executed in a tightly synchronized manner, optimizing their setup for a modest but achievable accuracy of over 90% within just four minutes. This project highlights the potential for out-of-the-box thinking in the AI/ML community, emphasizing innovation under constraints. Faced with challenges like camera calibration, color detection, and bandwidth limitations, the team successfully managed to operate a lightweight convolutional neural network with 670 parameters, showcasing a novel bandwidth setup that achieved around 30 KB/s. Their endeavor not only illustrates the flexibility of distributed training concepts but also invites further exploration into unconventional communication methods in AI, demonstrating both the feasibility and charm of such whimsical experiments in advancing machine learning techniques.