Visual-Steering-For-Deep-Neural-Networks

Abstract

This repository contains the source code to visually steer the CNN models into alligning with the human knowledge fed into them. This method is provided for integrating human knowledge with the convolutional neural network to improve its performance, reduce the biases that arise, and leverage human experience. [paper]

Citation

If you find this repository useful, please cite the following reference.

@InProceedings{10.1007/978-3-031-43247-7_4,
author="Elbeialy, Hossam and Ghaith, Arwa and ElBaradei, Mai H. and Alkhodary, Mohamed and Eissa, Alaa M. and Farag, Esraa A. and Fdawey, Omnia and Emad, Noran and Elshenawy, Ahmed and Shabaan, Mariam and Gamal, Marwa",
editor="Hassanien, AboulElla and Rizk, Rawya Y. and Pamucar, Dragan and Darwish, Ashraf and Chang, Kuo-Chi",
title="Visual Steering for Deep Neural Networks Using Explainable Artificial Intelligence",
booktitle="Proceedings of the 9th International Conference on Advanced Intelligent Systems and Informatics 2023",
year="2023",
publisher="Springer Nature Switzerland",
address="Cham",
pages="43--52",
abstract="Despite their remarkable revolution in artificial intelligence and promising performance, deep neural networks are susceptible to biases due to imbalanced training data or the way that training data is represented as features for the model; thus, they need a high level of transparency in critical fields, especially in the medical sector. Several explainability methods have been developed, aiming at enhancing the transparency of the model and increasing its trustworthiness. One of these methods is Grad-CAM, which can highlight the significant parts of an input that most influence the model's prediction and detect biases in the dataset. In this paper, a method is provided for integrating human knowledge with the deep neural network to improve its performance, reduce the biases that arise, and leverage human experience. This can be done via the attention maps generated by Grad-CAM, which are then edited by domain experts. The proposed method enables the edited attention maps to contribute to the training of the model and updating its parameters, leading to better generalisation and fewer inductive biases. This approach is applied to several image classification datasets: Imagenette2, BUSI, and skin cancer ISIC. Experimental results show that the generated attention maps highly correspond to the edited ones while having high classification performance.",
isbn="978-3-031-43247-7"
}

Methodology

The proposed approach is divided into 3 stages: explaining the model, manually editing the attention maps, and embedding human knowledge using a specific loss. First, Grad-CAM was used to generate attention maps describing the importance values of each image region according to the model to make a specific decision. These attention maps were then annotated by domain experts to represent what they believed were the image regions important to make that decision. Finally, the annotated attention maps were used to steer the model into aligning its behaviour with the human behaviour hence, embedding the human knowledge into the learning process of the model.

Enviroment

to install the required packages run the following command:

pip install -r requirements.txt

Execution

to run the code, you need to run the following command:

python Training.py [arguments]

Model-related Arguments

Argument	Description	Default Value
-m, --model	Model to train, check parameters.py for the list of supported models
--pretrained	Use pretrained model's weights from torchvision	True
--pretrained_weights_path	Path to pretrained model's weights
--model_path	Path to save the trained model (full model not the model weights)
--gamma	Gamma value of attention loss	10

Training-related Arguments

Argument	Description	Default Value
-b, --batch_size	Batch size for training	32
-e, --epochs	Number of epochs for training	100
--lr	Learning rate for training	0.001
-s, --seed	Random seed for reproducibility	42
-l, --loss	Loss function to use, check parameters.py for the list of supported loss functions
-o, --optimizer	Optimizer to use, check parameters.py for the list of supported optimizers
--class_weights	Use class weights to balance the loss function

Data-related Arguments

Argument	Description	Default Value
--image_folder	Path to the folder containing the images	imagenette2/Images/Train
--masks_folder	Path to the folder containing the masks	imagenette2/Masks/Train
--image_size	Size of the images	(224,224)
--test_image_folder	Path to the folder containing the test images	imagenette2All/val
-n, --num_classes	Number of classes in the dataset	10

Optimizer-related Arguments

Argument	Description	Default Value
--momentum	Momentum value of optimizer for SGD and RMSprop	0.9
--weight_decay	Weight decay value of optimizer	5e-4

Other Arguments

Argument	Description	Default Value
--out_path	Path to save the model	results
--gpu_id	Id of the GPU to use	0

Datasets

We applied this approach on 3 datasets:

• Imagenette2
• Breast Cancer Ultrasound Images (BUSI)
• International Skin Imaging Collaboration (ISIC)

Models

We applied this approach on several models from the ResNet family:

• ResNet50
• ResNet101
• ResNet152

Comparison of Accuracy, AUC, and Qualitative assessments of attention maps for Imagenette2 Dataset while using the annotation of 3948 images.

Approach	Model	Accuracy	AUC	Qualitative Measure
Without	ResNet50	99.49%	0.9999	3.4084
	ResNet101	97.73%	0.9995	3.5392
	ResNet152	98.22%	0.9997	3.2401
ABN	ResNet50	97.91%	0.9994	2.3048
	ResNet101	96.46%	0.9982	2.7847
	ResNet152	98.32%	0.9997	2.4483
Proposed Approach	ResNet50	97.58%	0.9995	1.8623
	ResNet101	98.11%	0.9996	3.1043
	ResNet152	97.89%	0.9991	2.4673

Comparison of F2 score and Qualitative assessments of attention maps for BUSI dataset while using the annotation of the whole dataset.

Approach	Model	F2 Score	Qualitative Measure
Without	ResNet50	0.8298	1.6049
	ResNet101	0.7736	1.8655
	ResNet152	0.7860	1.3899
ABN	ResNet50	0.8238	2.0944
	ResNet101	0.8066	1.9425
	ResNet152	0.7720	2.4787
Proposed Approach	ResNet50	0.8599	1.3292
	ResNet101	0.8429	1.4367
	ResNet152	0.8225	1.3881

Comparison of F2 score and Qualitative assessments of attention maps for ISIC dataset while using the annotation of the whole dataset.

Approach	Model	F2 Score	Qualitative Measure
Without	ResNet50	0.7061	2.9413
	ResNet101	0.5715	3.5000
	ResNet152	0.7156	2.9188
ABN	ResNet50	0.7207	2.2736
	ResNet101	0.6137	2.5259
	ResNet152	0.6961	2.2733
Proposed Approach	ResNet50	0.7405	1.6235
	ResNet101	0.7014	2.8837
	ResNet152	0.7126	2.8722