This document summarizes a final presentation on Powerpropagation, a new weight parameterization technique for neural networks that inherently leads to sparse models. Powerpropagation works by raising weights to a power α during training, which encourages "rich get richer" dynamics and sparsity. It was shown to achieve good accuracy on MNIST and Fashion MNIST datasets with only 10% remaining weights when α was set to 3. Powerpropagation can readily be combined with other techniques and is well-suited for continual learning tasks by overcoming catastrophic forgetting through gradient sparsity.