[Paper Review] Multifaceted Feature Visualization: Uncovering the Different Types of Features Learned By Each Neuron in Deep Neural Networks

Anh Mai Nguyen, Jason Yosinski+1 more|arXiv (Cornell University)|Feb 11, 2016

Cell Image Analysis Techniques22 references168 citations

TL;DR

The paper introduces Multifaceted Feature Visualization (MFV), an activation maximization technique that uncovers multiple facets of each neuron by initializing optimization from facet-specific mean images, improving image quality and revealing diverse neuron representations.

ABSTRACT

We can better understand deep neural networks by identifying which features each of their neurons have learned to detect. To do so, researchers have created Deep Visualization techniques including activation maximization, which synthetically generates inputs (e.g. images) that maximally activate each neuron. A limitation of current techniques is that they assume each neuron detects only one type of feature, but we know that neurons can be multifaceted, in that they fire in response to many different types of features: for example, a grocery store class neuron must activate either for rows of produce or for a storefront. Previous activation maximization techniques constructed images without regard for the multiple different facets of a neuron, creating inappropriate mixes of colors, parts of objects, scales, orientations, etc. Here, we introduce an algorithm that explicitly uncovers the multiple facets of each neuron by producing a synthetic visualization of each of the types of images that activate a neuron. We also introduce regularization methods that produce state-of-the-art results in terms of the interpretability of images obtained by activation maximization. By separately synthesizing each type of image a neuron fires in response to, the visualizations have more appropriate colors and coherent global structure. Multifaceted feature visualization thus provides a clearer and more comprehensive description of the role of each neuron.

Motivation & Objective

• Motivate the need to understand that neurons can be multifaceted and fire to diverse image types.
• Propose a method to visualize multiple facets of a neuron to better describe its role.
• Improve the quality and interpretability of activation maximization visuals.
• Demonstrate that higher-level neurons exhibit greater facet diversity than lower-level ones.
• Show that regularization and facet-specific initialization yield coherent global structure in visuals.

Proposed method

• Introduce Multifaceted Feature Visualization (MFV) to visualize multiple facets of a neuron.
• Cluster training set images of a class in a learned feature space to identify facets, using PCA, t-SNE, and k-means.
• Initialize activation maximization for each facet with the mean image from that facet’s cluster.
• Apply center-biased regularization to produce visually coherent, centralized objects.
• Compare MFV visualizations to prior activation maximization methods, highlighting color realism and global structure.

Experimental results

Research questions

• RQ1 Do neurons in DNNs have multiple distinct facets of response, and how do these facets vary across layers?
• RQ2 Can facet-wise initialization and regularization improve the interpretability and realism of activation-maximized images?
• RQ3 How do higher-layer neurons differ from lower-layer neurons in facet diversity and object composition?

Key findings

• Neurons are multifaceted feature detectors across layers, with higher-level neurons showing more diverse facets.
• MFV produces activation-maximized images with more natural colors and coherent global structure by visualizing each facet separately.
• Center-biased regularization reduces repeated fragments and centers the object, improving image quality.
• Visualization of hidden units shows increasing facet diversity starting from conv3, with fully connected layers often combining high-level concepts.
• Facets of a class can correspond to different real-world subtypes or perspectives, revealing rich internal representations.
• MFV outperforms previous activation maximization methods in terms of realism and interpretability of the synthesized images.