Natural Language Descriptions of Deep Visual Features

Abstract: Some neurons in deep networks specialize in recognizing highly specific perceptual, structural, or semantic features of inputs. In computer vision, techniques exist for identifying neurons that respond to individual concept categories like colors, textures, and object classes. But these techniques are limited in scope, labeling only a small subset of neurons and behaviors in any network. Is a richer characterization of neuron-level computation possible? We introduce a procedure (called MILAN, for mutual-information-guided linguistic annotation of neurons) that automatically labels neurons with open-ended, compositional, natural language descriptions. Given a neuron, MILAN generates a description by searching for a natural language string that maximizes pointwise mutual information with the image regions in which the neuron is active. MILAN produces fine-grained descriptions that capture categorical, relational, and logical structure in learned features. These descriptions obtain high agreement with human-generated feature descriptions across a diverse set of model architectures and tasks, and can aid in understanding and controlling learned models. We highlight three applications of natural language neuron descriptions. First, we use MILAN for analysis, characterizing the distribution and importance of neurons selective for attribute, category, and relational information in vision models. Second, we use MILAN for auditing, surfacing neurons sensitive to human faces in datasets designed to obscure them. Finally, we use MILAN for editing, improving robustness in an image classifier by deleting neurons sensitive to text features spuriously correlated with class labels.

• The paper introduces milan, an innovative method that uses pointwise mutual information to automatically generate detailed natural language descriptions of neuron activations.
• It employs a model-agnostic framework applicable to various architectures including CNNs, vision transformers, and generative models like BigGAN.
• The study demonstrates practical applications in analysis, auditing privacy-sensitive features, and editing spurious correlations to enhance model robustness.

Natural Language Descriptions of Deep Visual Features

The paper, authored by Hernandez et al., introduces a novel approach for interpreting the behavior of individual neurons in deep neural networks by automatically generating natural language descriptions that reflect neuron activation patterns on image data. This approach, termed milan, leverages pointwise mutual information to produce fine-grained, open-ended descriptions that could potentially enhance model interpretability and offer insights into the compositional and logical structure of learned features.

Key Contributions

1. Automated Neuron Descriptions: Unlike traditional methods which rely on fixed, predefined labels that only cover a subset of neurons, milan offers a flexible approach that labels neurons with descriptive language capturing nuanced attributes and structures identified in image data. By maximizing mutual information between linguistic descriptions and neuron activations, milan is capable of providing more detailed and contextually meaningful annotations.
2. Model-Agnostic Approach: Milan's labeling process is largely independent of specific model architectures and tasks, rendering it applicable across convolutional networks (CNNs) and vision transformers (ViTs). This adaptability was demonstrated in experiments spanning classifiers like AlexNet and ResNet152, as well as generative models like BigGAN and unsupervised models such as DINO.
3. Dataset Development: To train milan, the authors compiled the milannotations dataset, consisting of human-generated captions for neuron exemplar sets derived from several state-of-the-art vision models. This data grounds the mutual information calculation between neuron activations and natural language descriptions.
4. Applications: Three primary applications are explored. First, in analysis, milan's descriptions help ascertain the functional significance of different neuron classes, highlighting neurons with adjectives as crucial for performance. In auditing, milan identifies neurons that activate on human faces—even in models trained on anonymized datasets—demonstrating its utility for privacy assessment. Finally, in editing, milan facilitates the removal of neurons related to spurious correlations, thereby bolstering model robustness against adversarial attacks embedded in text features.

Implications and Future Directions

Practically, milan's ability to interpret neuron activations using natural language could streamline processes where understanding model rationales is crucial, such as in neural architecture search, model debugging, and increasing transparency in AI systems. Theoretically, this work underscores the potential to bridge the gap between human linguistic intuition and machine-learned representations, particularly within the context of vision tasks.

As future developments unfold, extending milan to cover more domains beyond vision, such as natural language processing or multimodal systems, could yield a more generalized framework for interpreting deep networks. Further, exploring the degree to which milan's interpretations can influence model design, tuning, and validation phases would be worthwhile, especially in dynamically shifting environments or under distributional shifts.