What Do Self-Supervised Vision Transformers Learn?

[arXiv, poster, slide]

Namuk Park^1*, Wonjae Kim², Byeongho Heo², Taekyung Kim², Sangdoo Yun²

¹ _{Prescient Design, Genentech} ² _{NAVER AI LAB} ^* _{Works done while at NAVER AI Lab}

This repository provides a PyTorch implementation of "What Do Self-Supervised Vision Transformers Learn (ICLR 2023)." In this paper, we delve into the properties of contrastive learning (CL) and masked image modeling (MIM) methods by comparing their learning patterns, providing valuable insights into the unique and complementary aspects of each approach. Specifically, we demonstrate the following:

1. CL primarily captures global patterns, while MIM focuses on local patterns;
2. CL is more shape-oriented, while MIM is more texture-oriented; and
3. CL plays a key role in the later layers of the model, while MIM focuses on the early layers.

These findings consistently indicate that CL is an image-level approach, and MIM is a token-level approach. In other words, CL learns the global invariant semantics of images, such as object shapes, whereas MIM learns token similarity semantics discerning which tokens are similar and which are not. Since they capture different types of information, their behavior differs, and they can even complement each other.

Let's find the details below 👇

How Do Self-Attentions Behave?

The self-attentions of CL and MIM reveal strikingly different behaviors. CL's self-attention captures global information, yet it generates homogeneous attention maps for all query tokens and heads. This implies that it primarily focuses on global invariant aspects and object shapes, possibly sacrificing interaction diversity among tokens. Conversely, MIM's self-attention mainly focuses on local areas, adeptly capturing neighboring similar tokens. This suggests that MIM excels at discerning the similarities and differences between tokens, outshining CL in this regard.

Consequently, CL and MIM behave differently. CL excels at linear probing tasks due to its ability to capture shapes, which helps recognize objects and distinguish images. Although MIM preserves the local information, the correlation with objects or content might not be as strong as with shapes. Furthermore, MIM is more scalable than CL as it effectively leverages diverse representations. Lastly, CL may not be suitable for dense prediction tasks, as self-attention maps and token features are uniform with respect to their spatial coordinates.

How Are Representations Transformed?

Analyzing representations from a token-level perspective reveals distinct differences between CL and MIM. CL processes all tokens in unison, while MIM handles each token individually. Consistent with their performance in linear probing tasks, CL effectively separates images by capturing object shapes, even though the tokens form clusters for each image. In contrast, MIM's tokens are intermingled, suggesting that they can recognize individual tokens well but lack linear separability. Likewise, CL exhibits a texture bias, while MIM is more shape-biased. This means that CL is robust against small data perturbations.

Which Components Play an Important Role?

Later layers of CL and early layers of MIM play significant roles. Interestingly, this aligns with the general consensus that early layers typically capture low-level features, such as local patterns, high-frequency signals, and texture information, while later layers capture global patterns, low-frequency signals, and shape information.

Since early layers are crucial in MIM, separating decoders from backbones could be beneficial for fully utilizing their advantages. For this reason, MAE is not only efficient but also an effective method that fully leverages the benefits. MAE reconstructs blurred images as it explicitly decomposes and utilizes the properties of implicit decoders.

Based on these observations, we also discover that CL and MIM can complement each other, and even the simplest harmonization can capitalize on their strengths. Remarkably, the most basic hybrid models outperform those pre-trained with either CL or MIM in both fine-tuning and linear probing accuracy.

In summary, we demonstrate that CL and MIM are complementary in three aspects: self-attention, representation, and architecture. This implies that CL learns image-level invariant information, while MIM learns token-level similarity information, leading to differences in their behaviors.

Okay, so what? Which method is supposed to be better? To put it simply, MIM is the better choice if we have access to a lot of resources, such as large model sizes and long training times. If not, CL is a good alternative. The hybrid model performs best, but it's significantly more expensive.

Getting Started

The following packages are required:

• pytorch
• timm
• matplotlib
• scikit-learn
• fastai
• einops
• gdown
• yacs

Install all the dependencies by running the command pip install -r requirements.txt.

The pretrained models from MoCo-v3 and SimMIM are provided. The weights are automatically downloaded when you run the code.

Self-Attention Analysis

Please refer to the self_attention_analysis.ipynb (available on Colab notebook) for analyzing the self-attention behaviors. This notebook includes snippets for visualizing attention map, calculating attention distance, and measuring normalized mutual information. Run all cells to analyze the properties of self-attentions.

Representation Analysis

Please refer to the representation_analysis.ipynb (available on Colab notebook) for analyzing the representational behaviors. This notebook includes snippets for Fourier analysis and token-level t-SNE visualization. Run all cells to analyze the properties of representations.

Citation

If you find this useful, please consider citing 📑 the paper and starring 🌟 this repository. Please do not hesitate to contact Namuk Park (email: namuk.park@gmail.com, twitter: xxxnell) with any comments or feedback.

@inproceedings{park2023ssl,
  title={What Do Self-Supervised Vision Transformers Learn?},
  author={Park, Namuk and Kim, Wonjae and Heo, Byeongho and Kim, Taekyung and Yun, Sangdoo},
  booktitle={International Conference on Learning Representations},
  year={2023}
}

Please also consider citing the following paper ("How Do Vision Transformers Work?" (ICLR 2022 Spotlight), code and summary ) if you find it useful for analysis techniques, such as Fourier analysis of representations.

@inproceedings{park2022how,
  title={How Do Vision Transformers Work?},
  author={Namuk Park and Songkuk Kim},
  booktitle={International Conference on Learning Representations},
  year={2022}
}

License

This project is distributed under MIT license,
except models/moco.py which is adopted from https://github.com/facebookresearch/moco-v3 under CC BY-NC 4.0.

Copyright 2023-present NAVER Cloud Corp.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.