MULTIRESOLUTION HASH ENCODING FOR NEURAL NETWORKS

Smart overview of the Invention

The patent application describes a method to enhance neural network performance by encoding inputs into a higher dimensional space using a hash function. This approach involves quantizing input coordinates, which represent points in a multi-dimensional space, and using these to generate vertex coordinates. These coordinates are then processed through a hash function to produce encoded coordinates, which are used to retrieve feature vectors from a learned table. This process ultimately improves training speed and accuracy of neural networks by providing a more efficient input encoding mechanism.

Background

Fully connected networks, also known as multi-layer perceptrons (MLPs), have become popular in computer graphics for tasks such as neural signed distance functions and light fields. These tasks benefit from encoding inputs into a higher-dimensional space to improve approximation quality. However, traditional methods can be costly in terms of training and evaluation. The described method addresses these issues by offering a more efficient and versatile encoding strategy that is applicable across various applications without requiring specific data structures.

Methodology

The method involves dividing the input space into axis-aligned voxels and using the vertex coordinates of these voxels as inputs to a hash function. The resulting encoded coordinates are used to access a table of learned feature vectors, which are then filtered based on the input point's coordinates. This encoded feature vector is fed into the neural network, rather than the raw input coordinates. The process is executed across multiple resolutions, allowing the system to capture both small and large features efficiently.

Advantages

The multiresolution hash encoding offers significant improvements in accuracy, versatility, and efficiency compared to conventional systems. By utilizing a versatile hashing scheme, the method can handle non-uniformly distributed input data, making it suitable for a wide range of applications. The use of multiple resolutions allows for the representation of diverse feature scales while maintaining a small memory footprint and computational cost. This approach also enables the neural network to resolve hash collisions internally, eliminating the need for explicit collision handling.

Applications and Implementation

The multiresolution hash encoding method is particularly beneficial for neural graphics primitive tasks, such as mapping 2D positions to 3D colors or predicting volumetric density and color. The implementation involves a small, learnable hash table that optimizes feature vectors during training. This architecture supports parallelization and fits well within processor caches, enhancing performance on graphics processors. The method allows for the use of smaller neural networks without compromising quality, significantly reducing computational requirements.