Reproduce “Feature-based Attention in Convolutional Neural Networks” 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
Grace W. Lindsay
The authors introduce feature-based attention (FBA), which biases a convolutional network toward the feature pattern associated with an object category. Its category-specific modulation is applied globally at a selected layer in one feedforward pass. On constructed four-image arrays, multiplicative, bidirectional FBA at layer 5 increased binary classification performance by 12.49 percentage points with category-specific tuning; this does not establish natural-scene localization.
Key insights
- Later representations, especially layer 5, benefit more consistently from feature-based modulation than earlier layers, aligning with evidence that attentional effects strengthen in higher visual areas.
- Multiplicative modulation generally outperforms additive modulation, and preserving negative feature-pattern values can help by suppressing features inconsistent with the attended category.
- Attention is category- and data-dependent: patterns estimated from as few as 20 examples can help, but categories with variable viewpoints or cluttered appearances produce less reliable modulation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.