Reproduce “Interpretable 3D Neural Object Volumes for Robust Conceptual Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nhi Pham, Artur Jesslen, Bernt Schiele, Adam Kortylewski, Jonas Fischer
CAVE, a framework developed at the Max Planck Institute for Informatics, unifies robustness to out-of-distribution data with inherent interpretability in deep neural networks. It leverages sparse 3D neural object volumes and weak 3D supervision to achieve competitive accuracy while providing faithful, spatially coherent concept explanations, quantified by a new 3D Consistency metric.
Key insights
- It is possible to unify OOD robustness and inherent interpretability within a single neural network architecture, overcoming the traditional trade-off.
- Sparse, high-level 3D concepts derived from neural object volumes can provide both efficiency and faithful, spatially coherent explanations comparable to denser representations.
- Objectively evaluating concept quality can be achieved through geometry-based consistency metrics, offering a more robust alternative to human-annotated parts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.