Top-down Neural Attention by Excitation Backprop

Jianming Zhang, Zhe Lin, Jonathan Brandt, Xiaohui Shen, Stan Sclaroff

TL;DR

This paper introduces Excitation Backprop, a biologically-inspired method to generate high-quality top-down attention maps from standard, pre-trained CNNs without requiring extra training or architectural changes. By formulating attention as a probabilistic Winner-Take-All process, the authors enable precise, weakly-supervised object localization and text-to-region association that outperforms existing visualization techniques.

What Problem Does This Paper Address?

While Convolutional Neural Networks (CNNs) excel at image classification, they are often treated as "black boxes." Understanding where a model looks to make a decision is crucial for both interpretability and downstream tasks like object localization.

Existing methods for visualizing attention often fall into two camps: they either produce noisy, pixel-level maps that lack semantic clarity (like simple gradient backprop), or they require specific architectural modifications (like Global Average Pooling for CAM). The authors sought a method that is biologically plausible, works with generic pre-trained models, and produces highly discriminative maps capable of localizing specific objects even in cluttered scenes.

Methods

The core innovation is Excitation Backprop (EB), which treats top-down attention as a probabilistic Winner-Take-All (WTA) process.

  1. Probabilistic WTA: Inspired by the biological "Selective Tuning" model, the authors model the top-down signal as a probability distribution. Instead of a hard "winner" neuron, they calculate the Marginal Winning Probability (MWP)—the likelihood that a specific neuron at an intermediate layer is responsible for the activation of the chosen output class.
  2. Excitatory Connections: The backpropagation scheme only passes signals through excitatory (positive) weights. This aligns with the intuition that top-down attention is looking for the presence of features that support the final classification.
  3. Contrastive Attention (c-MWP): To make the maps more discriminative, the authors introduce a contrastive version. By subtracting the attention map of a "dual" (negated) unit from the original map, they cancel out common background features and highlight only the neurons unique to the target class. This can be computed efficiently in a single backward pass.
Fig. 2. Deterministic WTA [3] vs. our probabilistic WTA for modeling top-down atten- tion. (a) Given a selected output unit, the red dots denote the winners identified by the top-down layer-wise deterministic WTA scheme in the processing cone, and the cyan ones are inhibited. (b) In our probabilistic WTA scheme, winner neurons are generated by a stochastic sampling process (shown in (c)). The top-down signal is specified by a probability distribution over the output units. The shading of a dot in (b) indicates the its relative likelihood of winning against the other ones in the same layer.

Datasets

The authors evaluated their method across several standard benchmarks and one massive new dataset:

  • PASCAL VOC07 & MS COCO: Used for the "Pointing Game," evaluating how accurately the attention peak hits the target object.
  • ImageNet: Used for dominant object bounding box localization.
  • Flickr30k Entities: Used to test text-to-region association (phrase localization).
  • Stock6M: A newly introduced dataset of 6 million weakly labeled images from a stock photo website, used to train a large-scale tag classifier for 18,000 different visual concepts.

Key Insights and Ideas

  • Interpretability as a Diagnostic Tool: The authors show that visualizing attention reveals what the model learned, not just where it looks. For example, in Figure 1, the model's attention for the tag "father" focuses on the child, suggesting the model learned a correlation between the presence of a child and the label "father."
  • Linearity is Key: Because the MWP is a linear function of the top-down signal, complex contrastive signals can be combined at the top layer before propagating downwards, making the process computationally efficient.
  • Layer Selection Trade-offs: There is a fundamental trade-off in CNN hierarchies: higher layers provide better semantic grouping but lower spatial resolution, while lower layers offer precise localization but less semantic relevance.
Fig. 1. A CNN classifier’s top-down attention maps generated by our Excitation Back- prop can localize common object categories, e.g. chair and glass, as well as fine- grained categories like boy, man and woman in this example image, which is resized to 224×224 for our method. The classifier used in this example is trained to predict ∼18K tags using only weakly labeled web images. Visualizing the classifier’s top-down atten- tion can also help interpret what has been learned by the classifier. For couple, we can tell that our classifier uses the two adults in the image as the evidence, while for father, it mostly concentrates on the child. This indicates that the classifier’s understanding of father may strongly relate to the presence of a child.

Key Results

  • The Pointing Game: On the difficult subsets of VOC07 and MS COCO, the contrastive MWP (c-MWP) outperformed the second-best method (like Grad or CAM) by approximately 10 percentage points using a GoogleNet architecture.
  • Dominant Object Localization: On ImageNet, the non-contrastive MWP achieved a localization error of 38.7%, which is competitive with state-of-the-art "Feedback Net" models that require 10-50 iterations of forward-backward passes.
  • Phrase Localization: On Flickr30k Entities, the method achieved competitive results against fully supervised baselines, particularly in rare categories like "animal" and "instrument," where human-annotated data is scarce.

Limitations and Future Work

  • Co-occurrence Bias: The model can struggle with objects that frequently appear together. For instance, it might falsely focus on a faucet when cued for a "sink" because the two are highly correlated in the training data.
  • Language Fusion: For the text-to-region task, the authors used a simple average of word-level attention maps. They suggest that integrating more advanced language models could better fuse these signals for complex phrases.
  • Resolution: While the method is efficient, the spatial resolution is still limited by the stride and receptive fields of the chosen intermediate layers.

Summary updated

Abstract

We aim to model the top-down attention of a Convolutional Neural Network (CNN) classifier for generating task-specific attention maps. Inspired by a top-down human visual attention model, we propose a new backpropagation scheme, called Excitation Backprop, to pass along top-down signals downwards in the network hierarchy via a probabilistic Winner-Take-All process. Furthermore, we introduce the concept of contrastive attention to make the top-down attention maps more discriminative. In experiments, we demonstrate the accuracy and generalizability of our method in weakly supervised localization tasks on the MS COCO, PASCAL VOC07 and ImageNet datasets. The usefulness of our method is further validated in the text-to-region association task. On the Flickr30k Entities dataset, we achieve promising performance in phrase localization by leveraging the top-down attention of a CNN model that has been trained on weakly labeled web images.
BibTeX
@article{zhang2016top,
  author = {Jianming Zhang and Zhe Lin and Jonathan Brandt and Xiaohui Shen and Stan Sclaroff},
  title = {{Top-down Neural Attention by Excitation Backprop}},
  journal = {Int. J. Comput. Vis.},
  year = {2016},
  doi = {10.1007/s11263-017-1059-x},
}
Figure