Reproduce “Deep Internal Learning: Deep Learning from a Single Input” 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

Tom Tirer, Raja Giryes, Se Young Chun, Yonina C. Eldar

This survey paper provides a comprehensive overview of deep internal learning techniques, which enable deep neural networks to learn effectively from a single input example by exploiting inherent data structures or adapting pre-trained models. The reviewed methods demonstrate improved robustness and generalization across various signal processing tasks, bypassing the need for large external datasets.

Key insights

  • Deep neural network architectures inherently act as effective implicit priors for natural signals, preferentially fitting desired structures before overfitting noise.
  • Internal statistical regularities, such as self-similarity within and across scales of a single image, can be effectively leveraged to learn complex tasks from limited data.
  • Combining pre-trained external knowledge with test-time internal adaptation yields superior performance and robustness against domain shifts and unknown degradation models.
  • Statistical loss functions like SURE can stabilize optimization in internal learning, reducing the reliance on precise early stopping and improving solution reliability.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.