Learned Initializations for Optimizing Coordinate-Based Neural Representations
This work offers an incremental improvement in efficiency and generalization for researchers and practitioners working with coordinate-based neural representations, particularly in scenarios with limited data.
This paper addresses the inefficiency of optimizing coordinate-based neural representations from random initializations for each new signal. By applying meta-learning to learn initial weight parameters for specific signal classes, the authors achieve faster convergence and improved generalization when only partial observations are available.
Coordinate-based neural representations have shown significant promise as an alternative to discrete, array-based representations for complex low dimensional signals. However, optimizing a coordinate-based network from randomly initialized weights for each new signal is inefficient. We propose applying standard meta-learning algorithms to learn the initial weight parameters for these fully-connected networks based on the underlying class of signals being represented (e.g., images of faces or 3D models of chairs). Despite requiring only a minor change in implementation, using these learned initial weights enables faster convergence during optimization and can serve as a strong prior over the signal class being modeled, resulting in better generalization when only partial observations of a given signal are available. We explore these benefits across a variety of tasks, including representing 2D images, reconstructing CT scans, and recovering 3D shapes and scenes from 2D image observations.