XCNN: Embedded Encoder-Decoder in Convolutional Networks Towards Explainable AI

This paper proposes a new explainable convolutional neural network (XCNN) which represents important and driving visual features of stimuli in an end-to-end model architecture. This network employs encoder-decoder neural networks in a CNN architecture to represent regions of interest in an image based on its category. The proposed model is trained without localization labels and generates a heat-map as part of the network architecture without extra post-processing steps. The experimental results on the CIFAR-10, Tiny ImageNet, and MNIST datasets showed the success of our algorithm (XCNN) to make CNNs explainable while offering a simple architecture that can be reapplied to any CNN classifier.

Implementation Setup:

Main libraries

Python (python=3.7)
PyTorch (torch==1.4.0)
Keras (keras==2.2.4)
Other required libraries (such as cv2, matplotlib, numpy, etc. used in the Training and Test codes)

Files and Folders:

1- Code: All codes are here

2- Models: Includes trained PyTorch XCNN models for heatmap generation; and Trained Keras VGG-16 models for the Innvestigate Tool

3- TestImages: Sample images used in the paper

4- tinyimagenet_data, mnist_data, cifar_data: These are the folders that will be used when you download corresponding datasets.

5- Results.ipynb: Regenerating the results shown in the paper and to generate new results using new images

Data Collection:

MNIST: Download in the code using Torchvision

CIFAR-10: Download in the code using Torchvision

TinyImageNet Download from https://tiny-imagenet.herokuapp.com/ in the 'tinyimagenet_data' folder

Training

System Used: Linux Ubuntu, GPU 1080 (8G), Cuda 10

Example: CIFAR-10 training

python TrainCifar.py -epoch 300 -lr 0.001 -wd 0.000001 -beta 0.9 -cuda -gpu 0 -batch 128 -lr_decay 0.95

XCNN model architecture can be changed in the Train files by changing the generator and discriminator kernels, fully connected layers, and other hyperparameters.
The XCNN model details can be found in XAI.py
Data.py prepares data for TinyImageNet
Default hyperparameters are the ones used in the paper

Testing/Demo

Please see the detail in 'Results.ipynb' to generate heatmaps by the XCNN and other methods. Example:

Paper/Citation:

Amirhossein Tavanaei, 'Embedded Encoder-Decoder in Convolutional Networks Towards Explainable AI', Arxiv 2020, https://arxiv.org/abs/2007.06712

tavanaei / ExplainableCNN