Tech Stack
Unity
C#
Description
A feedforward neural network built without a library — neuron and layer structures, sigmoid activation, and backpropagation all written directly in C#.
Most of the real work turned out to be preprocessing. A raw mouse path is noisy and arbitrary in length; smoothing and normalizing it into a fixed 12-vector representation is what made 16 gesture classes separable at all.
This is classical machine learning, not generative AI — worth stating plainly, because the two get conflated.
- Implemented neuron/layer structures, sigmoid activation, and backpropagation from scratch.
- Designed a 12-vector input representation with mouse-path smoothing and normalization.
- Supported 16 gesture patterns with a live classification UI.
Page Info
Network & preprocessing
Neuron and layer structures, sigmoid activation, backpropagation, and a 12-vector gesture representation.