Jul 2024/AI/ML/course
Paper Rock Scissors Classification
Many machine learning engineers throw massive models at simple problems to mask bad data. I built this image classifier to prove the exact opposite. The goal was to show that disciplined data augmentation and strict validation tracking generate far more value than unnecessary system complexity.
I designed a compact TensorFlow pipeline to classify hand gestures into paper, rock, or scissors. Instead of building a deeper neural network, I focused entirely on the preprocessing layer. By applying aggressive geometric augmentations to the dataset, the model was forced to learn true rotational invariance rather than memorizing the training set.
The biggest takeaway is that a model is only as robust as its validation discipline. Tracking the loss curves across epochs revealed exactly when the network started to generalize. This project reinforced that foundational computer vision is not about sheer compute power, but about controlling the learning environment to guarantee an inference-ready outcome.