TL;DR
This paper proposes two statistical methods—Method of Variances and Method of Relative Activities—to identify which specific nodes in a Deep Belief Network (DBN) represent relevant features versus "noise." By identifying and silencing nodes that represent irrelevant data (like a digit drawn over a face), the authors demonstrate significantly cleaner image reconstruction.
The Transparency Problem in DBNs
Deep Belief Networks (DBNs) are powerful tools for automated feature extraction, but they suffer from a "transparency" problem. When a DBN learns to represent data, it doesn't come with labels explaining what each node in its final layer actually sees.
In a real-world scenario like face recognition, an image might contain "side information" or noise—for example, a handwritten digit overlapping a face. A standard DBN will extract features for both the face and the digit. Without knowing which nodes represent the face and which represent the digit, it is difficult to filter out the noise to improve the accuracy of the primary task. This paper seeks to bridge that gap by providing a way to distinguish between task-relevant and task-irrelevant nodes.
Methods for Feature Distinction
The authors introduce two primary techniques to map the "functional geography" of the DBN's top layer:
1. Method of Variances
This approach assumes that specific features activate specific nodes. By feeding the network a group of images containing only one type of feature (e.g., only faces), the nodes responsible for representing those features will show significantly higher variance in their activation levels compared to others. A threshold (Variance > 0.1) is used to flag these "specialist" nodes.
2. Method of Relative Activities
This is a differential analysis technique. The authors compare the activations of the top layer for two different inputs: a "clean" image (e.g., just a digit) and a "composite" image (e.g., a face corrupted by that same digit). By subtracting the activations of the clean image from the composite, the "relative activity" reveals which nodes are specifically dependent on the additional feature (the face).
The "Inactivation" Trick
Once the "noise" nodes (those representing the digits) are identified, the authors "inactivate" them during the reconstruction process. Instead of just turning them off (setting to zero), they replace the node values with a "neutral value"—the average activation of those nodes when only faces are presented to the network. This maintains the statistical balance of the network's internal state.
Datasets and Architecture
The study utilized three datasets to train and test their 4-layer DBN (architecture: 2000-1000-500-100 nodes):
- CMU PIE face database: 10,000 face images used as the primary feature of interest.
- MNIST dataset: 5,000 handwritten digits used as the "noise" or secondary feature.
- Synthetic Corrupted Dataset: 5,000 images created by overlaying MNIST digits onto CMU PIE faces to test the separation methods.
Key Insights and Results
- Feature Partitioning: The most significant insight is that DBNs naturally partition different types of features into distinct groups of nodes in the final layer. They don't necessarily "mush" everything together; the features are there, just unlabeled.
- Neutrality over Zeroing: When trying to remove a feature (like a digit) from a reconstruction, simply setting the nodes to zero isn't as effective as setting them to a "neutral" average value.
- Qualitative Improvement: The visual results are striking. When reconstructing a face from a corrupted image (Face + Digit), the standard DBN output is blurry and retains the ghost of the digit. However, when the identified "digit nodes" are inactivated using the proposed methods, the reconstructed face is much clearer and the digit is effectively erased.
- Thresholding: The authors found that a variance threshold of 0.1 and a relative activity threshold of 0.7 were effective for their specific architecture in separating these features.
Limitations and Future Work
- Manual Thresholding: The methods rely on manually set thresholds (0.1 for variance, 0.7 for activity), which might change depending on the dataset or network architecture.
- Requirement for "Clean" Samples: To identify the nodes, the authors needed access to clean samples of the individual features (pure faces and pure digits). In a completely unsupervised real-world scenario where features are always mixed, these methods might be harder to apply.
- Scalability: The paper focuses on a two-feature problem (faces vs. digits). It remains to be seen how well these methods scale when trying to distinguish between dozens of overlapping feature types.
Summary updated