Reproduce “Convolutional Neural Networks Are Not Invariant to Translation, but They Can Learn to Be” 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
Valerio Biscione, Jeffrey S. Bowers
Convolutional Neural Networks are not inherently translation invariant; instead, they acquire this property through exposure to translated objects during training. This learned invariance generalizes to novel objects and datasets, but subsequent training on location-specific tasks can disrupt it due to catastrophic interference.
Key insights
- Most CNNs lack architectural translation invariance; their ability to generalize to novel locations is learned.
- Simple datasets with translated objects are sufficient to induce generalizable translation invariance in CNNs.
- Learned translation invariance can be disrupted by catastrophic interference when CNNs are fine-tuned on location-specific tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.