Vision Transformers Exhibit Artifacts in Feature Maps and Require Registers to Mitigate Them

Core Concepts

Large, sufficiently trained vision transformer models learn to repurpose low-informative tokens to store and process global image information, leading to artifacts in feature maps. Adding additional "register" tokens to the input sequence allows the model to isolate this behavior, resulting in smoother feature maps and improved performance on dense prediction tasks and unsupervised object discovery.

Abstract

The paper investigates artifacts observed in the feature maps of modern vision transformer models, including supervised (DeiT-III), text-supervised (OpenCLIP), and self-supervised (DINOv2) models. The authors find that these artifacts correspond to high-norm tokens that appear during inference, primarily in low-informative background areas of images, and are repurposed by the model for internal computations.
The authors propose a simple solution to this issue - adding additional "register" tokens to the input sequence of the Vision Transformer. These registers allow the model to isolate the behavior of repurposing low-informative tokens, resulting in smoother feature maps and attention maps.
The authors show that this solution fixes the artifact problem entirely for both supervised and self-supervised models, sets a new state of the art for self-supervised visual models on dense visual prediction tasks, enables object discovery methods with larger models, and leads to smoother feature maps and attention maps for downstream visual processing.
The key insights are:

Large, sufficiently trained vision transformer models learn to recognize redundant tokens and repurpose them to store and process global image information.
This behavior leads to artifacts in the feature maps, as the model discards local patch information in favor of global information.
Adding additional "register" tokens to the input sequence allows the model to isolate this behavior, resulting in smoother feature maps and improved performance.
The authors observe that the high-norm outlier tokens appear around the middle layers of the vision transformer, and only after a sufficiently long training of a sufficiently big transformer.
Linear probing experiments show that the outlier tokens hold less information about their original position in the image or the original pixels in their patch, but more global information about the image.

Stats

"a small fraction of the total sequence (around 2%)" of tokens have roughly 10x higher norm than the others.
These high-norm tokens appear around the middle layers of the vision transformer.
These high-norm tokens only appear after a sufficiently long training of a sufficiently big transformer.

Quotes

"the model discards the local information contained in these patches during inference."
"the model learns to recognize patches containing little useful information, and recycle the corresponding tokens to aggregate global image information while discarding spatial information."

Key Insights Distilled From

Vision Transformers Need Registers

by Timo... at arxiv.org 04-15-2024

https://arxiv.org/pdf/2309.16588.pdf

Deeper Inquiries

How do the register tokens learn to specialize and focus on different parts of the image, as observed in the qualitative analysis?

In the qualitative analysis, we observed that register tokens in the Vision Transformer models showed variability in their positional focus, with some registers focusing more on border areas while others focused on centered areas. This specialization and focus on different parts of the image can be attributed to the training process and the role of the register tokens in the model.

Training Signal: During training, the model learns to allocate attention to different parts of the image based on the task at hand. The register tokens, being additional learnable tokens in the sequence, adapt their attention patterns to capture specific information that is relevant for the model's performance.

Global Information: Register tokens, similar to the [CLS] token, are designed to capture global information about the image. This global context allows the model to understand the overall content and context of the image, leading to specialized attention patterns in different registers.

Specialization: As the model processes different images and tasks, the register tokens may specialize in focusing on specific features or regions of interest based on the patterns in the data. This specialization allows the model to extract and utilize relevant information effectively.

Diversity in Attention: The variability in attention patterns among register tokens may arise from the diversity of images in the dataset, task requirements, and the model's learning process. Each register may learn to focus on different aspects of the image to contribute to the overall representation.

Overall, the specialization and focus of register tokens on different parts of the image are a result of the model's training process, the role of registers in capturing global information, and the diversity of images and tasks encountered during training.

How do the insights from this work be applied to improve the interpretability and performance of other types of transformer-based models beyond vision transformers?

The insights from this work on register tokens and artifact removal in Vision Transformers can be applied to enhance the interpretability and performance of other transformer-based models in various domains beyond vision transformers. Here are some ways these insights can be leveraged:

Artifact Removal: The concept of using additional tokens like registers to mitigate artifacts and improve feature maps' quality can be extended to other transformer-based models. By introducing similar mechanisms to identify and address artifacts, models in NLP, speech recognition, and other domains can achieve smoother and more reliable representations.

Global Information Capture: The idea of using specialized tokens to capture global information can benefit models in NLP tasks such as language modeling, text generation, and machine translation. By incorporating register tokens to capture context and global dependencies, these models can improve performance on complex language tasks.

Interpretability: The focus on attention maps and token behavior can enhance the interpretability of transformer models in various domains. By analyzing the attention patterns of specialized tokens like registers, researchers can gain insights into how the model processes information and make decisions, leading to better interpretability and model understanding.

Performance Enhancement: By optimizing the training process to incorporate additional tokens for specific purposes, such as storing global information or addressing artifacts, transformer-based models in different domains can achieve improved performance on a wide range of tasks. This approach can lead to more robust and efficient models across various applications.

In conclusion, the insights from this work on register tokens and artifact removal can be generalized and applied to enhance interpretability and performance in diverse transformer-based models beyond vision transformers, offering new avenues for research and development in the field.