How Powerful Are Randomly Initialized Pointcloud Set Functions?
This work addresses the problem of understanding representation power in machine learning for researchers, revealing that random embeddings can be surprisingly effective, which is incremental as it builds on prior studies of untrained models.
The paper investigates the effectiveness of randomly initialized neural set functions for point cloud embeddings, finding that they can achieve close to or better accuracy than fully trained models in tasks like classification, with results showing linear separability and strong feature capture.
We study random embeddings produced by untrained neural set functions, and show that they are powerful representations which well capture the input features for downstream tasks such as classification, and are often linearly separable. We obtain surprising results that show that random set functions can often obtain close to or even better accuracy than fully trained models. We investigate factors that affect the representative power of such embeddings quantitatively and qualitatively.