Kilobyte Models Enable Ultra-Compact Neural Network Deployment

Traditional neural network deployment faces bottlenecks due to the large size of trained models, especially for over-the-air updates or on-device libraries. This research introduces an extreme compression method where the deployable artifact is not the full weights, but a compact "recipe" to regenerate them. The core idea builds on Mapping Networks, which represent network weights as a function of a small trainable latent vector and a fixed random basis. By storing only the latent vector and an integer seed (which can reproduce the basis and initialization), models can be reduced to kilobyte sizes. This approach achieves accuracy comparable to aggressively quantized models while requiring significantly less storage.