Top-Down Neural Attention by Excitation Backprop

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

Introduction

Convolutional Neural Networks (CNNs) have revolutionized computer vision, yet they are frequently criticized for being "black boxes." While a network might correctly classify an image as containing a "cat," understanding exactly which pixels or features led to that decision remains a challenge. This interpretability is not just a matter of curiosity; it is essential for debugging models, ensuring fairness, and enabling weakly supervised tasks like object localization where bounding box annotations are unavailable.

Existing visualization techniques, such as gradient-based saliency maps or deconvolutional networks, often struggle with noise or lack task-specificity. For instance, a generic attention map might highlight all salient objects in a scene, regardless of whether the user is inquiring about a "dog" or a "cat." This paper introduces Excitation Backprop, a biologically inspired framework that models top-down attention as a probabilistic process to generate highly discriminative, class-specific attention maps without requiring architectural changes or additional training.

Methodology: Excitation Backprop

The core of the proposed method is a novel backpropagation scheme inspired by the "Selective Tuning" model of human visual attention. The authors formulate top-down attention as a Marginal Winning Probability (MWP).

Probabilistic Winner-Take-All

Unlike traditional deterministic "Winner-Take-All" (WTA) mechanisms that produce binary outputs, Excitation Backprop treats the backward pass as a stochastic process. A connection between a "parent" neuron in a higher layer and a "child" neuron in a lower layer is considered "excitatory" if the weight is non-negative. The probability of a child neuron being responsible for the parent's activation is calculated based on the relative strength of the bottom-up activations and the excitatory weights. By propagating these probabilities backward layer-by-layer, the model produces a "soft" attention map that reflects the relative importance of different image regions for a specific high-level concept.

Contrastive Top-down Attention

A significant challenge in top-down attention is that different classes often share low-level features or background contexts. To isolate the features unique to a specific target, the authors introduce Contrastive MWP (c-MWP).

For any given target class (e.g., "man"), the method virtually constructs a dual "non-target" unit (e.g., "not-man"). By calculating the attention map for both the target and its negation, and then subtracting the latter from the former, the model effectively cancels out common activations. This "high-pass filter" effect significantly cleans the resulting maps, highlighting only the most discriminative regions for the specific query.

Task-specific attention maps generated by Excitation Backprop for various object categories and relationships in a single scene.

Experimental Evaluation

The authors evaluated Excitation Backprop across several benchmarks, focusing on its ability to localize objects using only image-level labels.

The Pointing Game

To rigorously test localization, the authors introduced "The Pointing Game." In this evaluation, a "hit" is recorded if the maximum point in the generated attention map falls within the ground-truth boundary of the object. On the PASCAL VOC07 and MS COCO datasets, c-MWP consistently outperformed existing baselines like Grad-CAM and Deconvolution. Notably, on the "difficult" subset of VOC07 (containing small or occluded objects), c-MWP achieved an accuracy of 72.3% using a GoogleNet architecture, compared to 61.4% for standard gradient-based methods.

Large-Scale Localization and Text-to-Region Association

The method's scalability was demonstrated using the Stock6M dataset, comprising 6 million weakly labeled images. The model successfully localized a massive vocabulary of approximately 18,000 tags, ranging from fine-grained species to specific body parts and actions. Furthermore, on the Flickr30k Entities dataset, the approach proved effective for text-to-region association, linking natural language phrases to specific image regions with performance comparable to fully supervised state-of-the-art models.

Discussion and Limitations

Excitation Backprop offers a powerful, "zero-retraining" solution for interpreting pre-trained CNNs. However, some limitations remain. The resolution of the maps is inherently tied to the layer from which backpropagation begins; higher layers offer better semantic context but coarser spatial resolution. Additionally, while the method is excellent at "pointing" to the most discriminative part of an object (like a bird's head), it does not always capture the full spatial extent of the object.

Future work may involve integrating more sophisticated language models to handle complex phrase localization and exploring ways to better capture the full geometry of objects while maintaining high discriminability.

Summary updated

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