狠狠撸

Trustworthy AI
LFI-CAM: Learning Feature Importance
for Better Visual Explanation
Kwang Hee Lee, Chaewon Park, Junghyun Oh, Nojun Kwak

Trustworthy AI
Overview
? We propose a novel architecture, LFI-CAM, which is trainable for image classification
and visual explanation in an end-to-end manner.
? LFI-CAM’s attention map generated during forward propagation, leverages to improve
the classification performance through the attention mechanism.
? Feature Importance Network (FIN) focuses on learning the feature importance
instead of directly learning the attention map to obtain a more reliable and consistent
attention map

Trustworthy AI
Related Works
Perturbation-based Method: LIME, RISE
Response-based Method: CAM
Gradient-based Method: Grad-CAM, Grad-CAM++ Hybrid-based Method: Score-CAM, ABN

Trustworthy AI
Related Works
? Class Activation Mapping

Trustworthy AI
Related Works
? Grad-CAM

Trustworthy AI
Related Works
? LIME (Local Interpretable Model-agnostic Explanations)
Ribeiro MT, Singh S, Guestrin C. " Why should i trust you?" Explaining the predictions of any classifier. InProceedings of
the 22nd ACM SIGKDD international conference on knowledge discovery and data mining 2016 Aug 13 (pp. 1135-1144).

Trustworthy AI
Related Works
? RISE (Randomized input sampling for explanation of black-box models)
Petsiuk V, Das A, Saenko K. Rise: Randomized input sampling for explanation of black-box models.
arXiv preprint arXiv:1806.07421. 2018 Jun 19.

Trustworthy AI
Related Works
? Score-CAM

Trustworthy AI
Related Works
? Attention Branch Network (ABN)

Trustworthy AI
Motivation
? Conventional CAM variants still have the limitations such as
classification performance degradation, multiple forward
computing, and additive backpropagation.
? In particular, ABN outputs unreliable and inconsistent
attention maps through several experiments with various
hyper-parameters.
? Although Score-CAM achieves better visual performance
with less noise and better stability than the gradient-based
approaches, multiple forward computing makes the
generation of visual explanation very slow.
? Inspired by ABN and Score-CAM, we proposed a new
architecture for image classification and visual explanation in
an end-to-end manner and outputs a more reliable and
consistent attention map.
ABN
ABN
LFI-CAM
LFI-CAM

Trustworthy AI
Proposed Method

Trustworthy AI
Feature Importance Network (Attention Branch)
? While ABN directly learns the class activation map
in the attention branch, we replaced the ABN
model’s attention branch with a new network
architecture, “Feature Importance Network (FIN)”.
? FIN helps the LFI-CAM model learn the feature
importance to generate better class activation map
than ABN.
? The attention map is generated by the weighted
sum of the feature maps from the last
convolutional layer and the learned feature
importance vector.
? To learn the feature importance, in a similar way to
the Score-CAM, we convert the original input into
a gray image, which is down-sampled to the size
of the feature map.

Trustworthy AI
Perception Branch
? The perception branch takes the original input
image as an input and outputs the final probability
of each class.
? The attention map is applied to the feature maps
by the attention mechanism.
? The attention mechanism helps the attention
map improve the classification performance.

Trustworthy AI
Training
? LFI-CAM model is trained in an end-to-end manner using training loss
calculated as the combination of the Softmax function and cross-
entropy at the perception branch in image classification task.
? The FIN is optimized by the attention mechanism of the perception
branch to improve the classification accuracy without any additional
loss function.

Trustworthy AI
Experiments
? Experimental Settings on Image Classification
? Datasets: CIFAR10, CIFAR100, STL10, Cat&Dog and ImageNet
? Baseline Models: CIFAR ResNet backbone (ResNet 20, 32, 44, 56, 110)
ImageNet ResNet backbone (ResNet 18, 34, 50, 101, 152)
? Optimizer and Hyper-parameters
? SGD with momentum
? 300 epochs for CIFAR10,100, STL10 and Cat&Dog
? 90 epochs for ImageNet
? Learning rate is initialized with 0.1 and later on divided by 10 at 50% and 75% of the total number of
training epochs
? Batch size of 128 for CIFAR ResNet backbone
? Batch size of 256 for ImageNet ResNet backbone

Trustworthy AI
Experiments- Visual Explanation Evaluation

Trustworthy AI
Experiments- Effectiveness of Feature Importance Network

Trustworthy AI
Experiments- Accuracy on Image Classification

Trustworthy AI
Experiments- Stability Evaluation of Visual Explanation

Trustworthy AI
Summary
? We proposed the LFI-CAM model, which is trainable for
image classification and produces better visual explanation in
an end-to-end manner.
? Feature Importance Network (FIN) helps our model focus on
learning the feature importance to generate a more stable
and reliable attention map.
? LFI-CAM is on par with ABN in terms of classification
accuracy and outmatches ABN in terms of attention map
quality.

狠狠撸

LFI-CAM: Learning Feature Importance for Better Visual Explanation

Recommended

More Related Content

What's hot (18)

Recently uploaded (20)

LFI-CAM: Learning Feature Importance for Better Visual Explanation