On visual explanation of supervised and self-supervised learning

This project is part of my MSc Thesis https://pergamos.lib.uoa.gr/uoa/dl/frontend/el/browse/3256221 for my postgraduate studies in Data Science and Information Technologies (NKUA).

The primary objective of this project is to interpret both supervised and self-supervised models, using either convolutional neural networks or visual transformers as a backbone. Variations of visualization methods are used, based on class activation maps (CAM) and attention mechanisms. Given an input image, these methods provide us with a saliency map that is used to interpret the network prediction. This map indicates the regions of the image that the model pays the most attention to. We evaluate these methods qualitatively and quantitatively. We further propose new alternative or complementary visualization methods, which show where important information can be hidden inside the network and how to reveal it. These new methods further improve the quantitative results. Our study highlights the importance of interpretability, shows some common properties and differences in the way supervised and self-supervised models make their predictions and provides valuable information on both the models and the visualization methods.

Backbone architectures

Self-supervised approaches

DINO
MoCo v3

The weights for ResNet-based and DeiT-based models trained with MoCo v3 approach on ImageNet can be found https://github.com/facebookresearch/moco-v3/blob/main/CONFIG.md.

Download them and put them into: /path/to/On_visual_explanation_of_supervised_and_self-supervised_learning/storage/pretrained_weights/

Visualization Methods

Methods based on CAM

Attention-based Methods

Raw attention maps from different layers
Rollout

Dataset

The dataset we use for the experiments is a subset of the validation set of ImageNet and can be found https://github.com/EliSchwartz/imagenet-sample-images.

Download it and put it into: /path/to/On_visual_explanation_of_supervised_and_self-supervised_learning/storage/imagenet_1000samples/

Run Locally

Clone the project

  git clone https://github.com/DimitrisReppas/On_visual_explanation_of_supervised_and_self-supervised_learning.git

Install dependencies

Install the requirements for this project

  pip install -r /path/to/On_visual_explanation_of_supervised_and_self-supervised_learning/storage/requirements.txt

Evaluate models based on ResNet

Example

Evaluate ResNet 50 qualitatively and quantitatively using GradCAM with:

  python Res_based_evaluation.py --use-cuda --method gradcam --arch res_50 --data_path ../storage/imagenet_1000samples/ --output_dir ../storage/quant_results/ --saliency True

Evaluate models based on DeiT

Example

Evaluate MoCo v3 that uses DeiT as backbone quantitatively using Rollout method with:

  python Deit_based_evaluation.py --use-cuda --method rollout --arch moco_v3_deit_base --data_path ../storage/imagenet_1000samples/ --output_dir ../storage/quant_results/ --saliency False

Results

Bellow, some qualitative and quantitative results are presented:

Saliency maps for a given input image, obtained from different models that are based on ResNet 50, when using four different CAM-based methods

Qualitative and quantitative results for all ResNet-based models, when using the GradCAM method

Quantitative evaluation of CAM-based saliency maps for models using ResNet as a backbone

Visualization of Raw attention maps derived from different layers and from models that use DeiT as a backbone

Qualitative results for DeiT-based models when the Rollout method and Raw attention maps, derived from the last layer, are used

Quantitative results for DeiT-based models when Raw attention maps, derived from the last layer, are used.