| Sammanfattning: | <p>In machine learning, neural networks are typically treated as models to be used for various tasks such as classification, regression or generative modeling. In this thesis, we take a different perspective and treat neural networks as data instead of models.</p>
<p>Indeed, in deep learning, data is often represented by arrays, such as a 2D grid of pixels for images. However, the underlying signal represented by these arrays is often continuous, such as the scene depicted in an image. A powerful continuous alternative to discrete arrays is then to represent such signals with an implicit neural representation (INR), a neural network trained to output the appropriate signal value for any input spatial location. An image for example, can be parameterized by a neural network mapping pixel locations to RGB values. In this thesis, we investigate the consequences and compelling properties of using such neural networks as data. We motivate and explore this approach through two main applications: compression, where we store data as neural networks, and generative modeling, where we train generative models directly on datasets of neural networks.</p>
<p>For data compression, we propose to store the quantized weights of an INR as a compressed code for a given signal (such as an image), instead of directly compressing the discretized signal. We show how this allows us to build a single neural codec that is seamlessly applicable to multiple data modalities, from images and audio to medical and climate data. For generative modeling, we propose to learn distributions of INRs, either implicitly by training on array data or explicitly by directly training models on datasets of INRs. We demonstrate how such an approach leads to compelling algorithms for a range of machine learning tasks, including 3D shape inference and generative modeling of neural radiance fields.</p>
|
|---|