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.
- 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.
- 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.
- 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.
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.
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