Reproduce “Totally Looks Like - How Humans Compare, Compared to Machines” 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
Amir Rosenfeld, Markus D. Solbach, John K. Tsotsos
York University researchers demonstrate that state-of-the-art deep convolutional neural networks are largely unable to reproduce the abstract image similarity judgments humans make on the challenging new "Totally-Looks-Like" dataset. Machines achieved a Recall@1 of less than 6% on this dataset, contrasting sharply with high human agreement.
Key insights
- Human perception of abstract image similarity is remarkably consistent and relies on diverse, high-level attributes not easily captured by current DCNN features.
- Existing machine-learned visual representations, primarily optimized for tasks like object classification, fail to generalize to the multi-faceted, high-level, and often abstract criteria humans use for 'in-the-wild' similarity.
- The TLL dataset provides a unique and challenging benchmark that effectively highlights a fundamental gap in current AI's ability to emulate human-like visual understanding beyond basic recognition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.