Convolutional Neural Networks by Tagging Filters

Abstract

Convolutional neural networks (CNNs) have achieved astonishing performance on various image classification tasks, but it is dificult for humans to understand how a classification comes about. Recent literature proposes methods to explain the classification process to humans. These focus mostly on visualizing feature maps and filter weights, which are not very intuitive for non-experts. In this paper, we propose FilTag, an approach to efectively explain CNNs even to non-experts. The idea is that if images of a class frequently activate a convolutional filter, that filter will be tagged with that class. Based on the tagging, individual image classifications can then be intuitively explained using the tags of the filters that the input image activates. Finally, we show that the tags are useful in analyzing classification errors caused by noisy input images and that the tags can be further processed by machines.

Keywords

• CNN
• images
• explainable AI
• semantic interpretability

-

dog, parrot, cat

… toucan, parrot … … parrot … …

… … … … parrot

most activated

filter parrot part shows a visual explanation. The lower part contains an image, but it does not explain the role of that channel across all possible input images. Hohman et al. [ 6 ] try to overcome this problem by aggregating particularly important neurons and identifying relations between them.

Other approaches focus on filters, the discerning feature

of CNNs. For example, Zeiler and Fergus [ 7 ] visualize the iflter weights to illustrate the patterns these filters detect.

However, these visualizations are based on the inputs

of the layers to which the respective filter belongs to.

Thus, only the filter patterns of the first layer can be directly associated with patterns on the input image of the network. To overcome this, the method Net2Vec [8] quan

1. Introduction

Deep convolutional neural networks (CNNs) are the state

of-the-art machine learning technique for image classification [1, 2]. In contrast to traditional feed-forward neural networks, CNNs have layers that perform a convolutional step (see Figure 2 for the relations in a convolution). Filters are used in a convolutional step which outputs a feature map in which activated neurons highlight certain patterns of the input image. Although CNNs achieve high accuracy on many classification tasks, these models do not provide an explanation (i.e., decisive information) of the classifications. Thus, researchers recently focused on methods to explain how CNNs classify images.

Related Work. Some of the earliest works on explaining

CNNs focus on visualizing the activations of individual neurons [3, 4]. However, these methods cannot explain as no human-understandable explanation is used. Olah et al. [ 5 ] defined a semantic dictionary by pairing every neuron activation with its abstract visualization using a channel attribution, determining how much each channel contributes to the classification result. This may explain the role of a channel in the classification of an individual USA (M. Färber) iflter embeddings. Alternatively, Network Dissection [ 9 ] uses human-labeled visual concepts to bring semantics to the convolutional layers. However, visualizations and embedding filters only explain the outcome of a model implicitly, whereas we assign explicit tags to filters which can be understood by non-experts. Most visualizations used for explaining CNNs are similar to the upper example in Figure 1, which visualizes the most activated convolutional filters. Clearly, such visualizations are dififcult to understand on their own. Adding an explicit explanation such as a semantic tag (e.g. “dog,” “parrot,” “cat,” or “toucan”) as shown in the bottom example would dramatically improve the explanation, including for nonexperts.

Contribution. Our contribution is threefold. First, we introduce FilTag, an automatic approach to explain the role of each convolutional filter of a CNN to non-expert humans. We use the fact that each filter is dedicated to a specific set of classes [ 7, 10, 11, 12 ]. Indeed, the idea of FilTag is to quantify how much a filter is dedicated to a class, and then tag each convolutional filter with a set of particularly important classes. The lower part of Figure 1 shows an example of what a CNN tagged in this way could look like. In that example, the rightmost iflter highlighted in red plays a role in classifying parrots, whereas the filter in the middle only plays a role in classifying birds in general, as both, toucans and parrots are both birds. This filter extracts features that are specific to these classes (e.g. wings, feathers, etc.). Second, our approach can also be used to explain the classification of an individual image. In the example in Figure 1, the classification of the input image as a parrot would be explained by the union of the tags of the activated filters, which FilTag is suitable to analyze classification errors. We analyze our approach with thorough experimentation using as a data set. All source code is available online.1 multiple CNNs, including VGG16, as well as ImageNet 2. Approach tions based on the role of each filter in a CNN (independent from concrete input images) using our concept of iflter tags. Then, in Section the filters that it activates. ular input image can be explained, namely in terms of

2.1. Explanations of Filters

Our explanation of filters works in two steps. In the first

step, we quantify how much each filter is activated by

1https://github.com/michaelfaerber/FilTag

are all animals, particularly tagged with parrot. Third, image, i.e. the output in the feature map for a given filter input image RGB channels * * and therefore use the most activated feature map while our approach focuses on a set of images of the same class.

Given a pre-trained CNN with a set of convolutional laydata set with labels ∈

from a set of labels , let with its respective set of filters (⋅) and a labeled be an input image and ∈

a convolutional layer. First, we collect the activations in the feature map to get the importance of the filters regarding an input (see terminology in Figure 2). Second, we scale these activations per layer between [0, 1]. In scaling the activations, we ensure that no image is overrepresented with overall high activation values. We scale the activations per layer because each layer has its specific pattern compositionality of filters. For example, the first convolutional layers detect simple patterns such as lines and edges whereas the layers, in the end, detect compositional objects [ 7 ]. Let (, , , )

be such a scaled activation in the th element in the feature map calculated from In Section 2.1, we propose a method to provide explana- structures which match better to human-understandable 2.2, we explain how a partic- image and filter ∈ in convolutional layer . In of (,̄ , ) over one class where | | is the number of as ConceptNet [ 13 ] or FAIRnets [ 14 ] can bring more inimages in class . This way, (, ) is the averaged value sights. ConceptNet is a semantic network with meanings of all activations of the images in one class respective its of words and FAIRnets is a neural network graph with iflter in layer . Thus, we can rank the classes according metadata about the architecture. For example, in Figure 1, to the highest averaged activation of the filter per layer if we input an image of a car but the most activated filters which will be the decisive criterion for the labeling. We, have tags of animals, we can conclude that the wrong therefore, compare the received values for each feature iflters were activated. map. We repeat these steps for all images in per label class.

Filter Tagging. We tag the filters according to their 3. Experiment corresponding values received in (, ) with the label of the input image class. We are interested in the feature 3.1. Experimental Setup maps with high activations of a certain class because they indicate important features associated with that class [ 6 ]. We define two methods to select those feature maps per class and per layer (because of the mentioned complexity in diferent layers): (i) -best-method (choose the feature maps with highest activation values) and (ii) -quantile-method (choose the -quantile of feature maps with highest activation values). These tags serve as an explanation of what the filter does. For example, in Figure 1, the leftmost activated filter has the three tags dog, parrot and cat, which suggests that this filter plays a role in recognizing animals.

Data Set. Following related work, we use ImageNet [ 16 ] from ILSVRC 2014 to conduct experiments on the introduced approach. This data set contains over one million images and 1, 000 possible class labels including animals, plants, and persons. Each class contains approximately 1, 200 images. We use a holdout split, using 80% of the images to tag the filters, while ensuring that there were at least 500 images from each class in the set, and the remaining 20% to test the explanations.

Baseline. We compare our approach with two state-ofthe-art visualization methods in explaining neural networks. The selection of the methods was based on their focus on feature visualization. One of the methods used 2.2. Explanations of Individual provided the fundamental basis of visualization of feaClassifications tures and uses minimal regularization [ 15 ], the other method uses optimization objectives [4].

While previous visual methods for explaining filters are Implementation. We implemented our method in dificult for humans to understand, textual assignment Python3 and used TensorFlow as deep learning library. can lead to unambiguous explanations (as later seen in The experiments were performed on a server with Inour experiments in Figure 3). To get an explanation given tel(R) Xeon(R) Gold 6142 CPU@2.60 GHz, 16 physical an input, we assume that the tags have a better informa- cores, 188GB RAM and GeForce GTX 1080 Ti. We used tion value with the classification of the CNN if the tags pre-trained neural network models from Keras Applimatch with the classification output. Therefore, we want cations. The filters of a VGG16 were explained in the to measure the hit of the prediction with the tags in the experiments using the introduced method. VGG16 was most activated filters. To do this, we determine the most used as CNN as it is frequently used in various computer frequently occurring labels for each image of a class ac- vision applications. We also evaluated on VGG19 and cording to the previous mentioned method using the InceptionNet but omit them due to page limitations. metric Hits@ . Hits@ measures how many positive label tags are ranked in the top- positions. For example, in 3.2. Analysis of the Explanations Figure 1, the classification of the input image as a parrot is explained by its high activation of filters tagged with parrot.

In this analysis, we want to study the explanations of

the filters using -best-method, with = 1 , in order to provide a better comparison with the state-of-the-art 2.3. Analysis of Classification Errors methods since they frequently visualize the most activated feature map. Figure 3 shows exemplary the visual FilTag can be used for error analysis using Hits@ . Tak- explanations of the baseline methods, and the tags of our ing misclassified input images, Hits@ indicates if the approach FilTag. As shown, the visual explanations of most relevant filters were activated. If Hits@ is high, we the baseline methods [ 15, 4 ] do not provide satisfactory can assume that there are similar features of the misclas- comprehension. At first sight, there is not much to undersified class and original image. Analyzing the tags, we stand. Considering our tags, one can imagine what the may find correlations in their semantics. Furthermore, visualizations display. We additionally include pictures linking the tags and filters to knowledge graphs such corresponding to our tags, to show the information value compared to only visualizations of the filters. Filter 95 seems to recognize a lampshade especially a trapezoidal 0.8 k=1 shape. Filter 150 is only tagged with cannon, i.e. the filter 0.7 kk==255 is specific for this class. Filter 288 detects a head of a 0.6 q=1% goldfinch especially with consideration of the yellow and q=5% black pattern. Filter 437 and Filter 462 recognize ears of sn@0.5 q=25% brown dogs and the body of snakes, respectively. This itH0.4 information would be hard to retrieve without the tags. 0.3 Even without considering the visualizations, one has a 0.2 good impression of what a filter detects. For example, it is quite impressive that Filter 288 detects this black yel- 0.1 0 10 20 30 40 50 low pattern which we can follow from the tags goldfinch, n toucan, and european fire salamander . As well, Filter 95 detects the trapezoid in table lamp, yurt, and lampshade. Figure 4: Hits@ with diferent and on ImageNet

In addition to comparing our method to the state-ofthe-art methods in CNN explanations, we linked the tags to concepts from ConceptNet [ 13 ] to achieve a coarsen- accuracy. If the labels, and thus Hits@ , do not correlate ing of common tags. ConceptNet is a semantic network with the output of the neural network, and thus with the with meanings of words. This comparison revealed that accuracy, then the filters have not been tagged sensibly many tags have both visual and semantic commonali- with our approach to gain an accurate explanation. We ties (e.g., see Filter 437 in Figure 3, rhodesian ridgeback, will interpret Hits@ and accuracy with diferent hyperbloodhound and redbone are all of type dog). Following parameters and , respectively. In Figure 4, we compute this evaluation process, we manually reviewed 100 filters Hits@ with the test set from ImageNet depending on in the context of common visual and semantic common- and . We can see that Hits@ increases for increasing , alities. Here we found 88% conformance with common and . For = 25% and = 50 , we even get a hit rate of tags in the filters. 80% over all 1,000 object classes. This result shows that FilTag can be taken as a significant explanation for the 3.3. Impact of Hyperparameters classification. For example, we have observed that the class shoji gets the highest hit rate of 98.47% followed by In the following we evaluate which impact the hyperpa- the classes slot, odometer and entertainment center with rameters and have on the correlation of Hits@ and also around 98%. This correlates with the likelihood of the best classes, which are exactly the same classes: shoji (81.22%), slot (92.30%), odometer (91.73%) and entertainment center (82.89%). Likewise, Hits@ also correlates with the accuracy of the worst classes, which are spatula, schipperke, reel, bucket, and hatchet. These results fit to the top-1 accuracy of VGG16 with 74, 4% for all classes.

The high correlation with Hits@ and accuracy shows that the relevant features, labeled by our approach, are in fact detected from the images, which confirms the hypothesis that the tags are useful to generate explanations by means of our approach. However, for larger values of we observed that the interpretability decreases because the number of tags increases for each filter. This makes it harder to find similarities between the classes.

Thus, there is a trade-of between expressiveness for the classification and interpretability for the filters.

3.4. Using the Explanations

(a) Mortarboard (b) Computer Figure 5: Example images from ImageNet neural network to assign it correctly. Moreover, it is an old computer, whereas the other images in ImageNet generally represent rather modern computers. In order to classify this image correctly, further images showing old computers from the side have to be included to change the distribution and train the VGG16 to classify this image correctly.

4. Conclusion

FilTag can be used for error analysis using Hits@ . Tak

ing misclassified input images, Hits@ indicates if the We have introduced FilTag, an approach to provide most relevant filters were activated. If Hits@ is high, human-understandable explanations of convolutional filwe can assume that there are similar features of the mis- ters and individual image classifications. These tags can classified class and original image. Analyzing the tags, be used to query and identify specific filters that are we may find correlations in their semantics. relevant for feature detection. In contrast to state-of

Figure 5 (a) shows an image of the class mortarboard the-art explanations, our approach allows for explicit, in ImageNet. Using VGG16, the class academic gown is non-visual explanations which are more understandable predicted with a confidence of 83.8%, while the actual for non-experts. class mortarboard is predicted with a confidence of only A limitation of our approach is the use of the class 16.2%. Considering the image, we notice that both ob- labels as tags to describe the filters. As a result, filters jects are part of this image, making this result reasonable. are not described in terms of specific objects such as ears, Reviewing the activated filters, we observe that filters wings, or legs. We would like to address this limitation tagged by FilTag with the tag mortarboard, as well as in the future by using ConceptNet and other knowledge with the tag academic gown, are usually activated. As bases to identify commonalities of the tags and thus add a result, we can verify that features are extracted from specific object descriptions to the filters. these two classes and used for prediction. This allows to give non-experts an understanding of the reason for the misclassification, as often features of the other class are References extracted from this image. Likewise, we can use the information to increase the number of images in which the [1] K. Simonyan, A. Zisserman, Very Deep Convolumortarboard is the actual class but not in the main focus tional Networks for Large-Scale Image Recognition, of the image, in order to continue learning the network in: 3rd International Conference on Learning Repto make the predictions more accurate. resentations, ICLR 2015, 2015.

Figure 5 (b) shows an image from the class computer. [2] C. Szegedy, V. Vanhoucke, S. Iofe, J. Shlens, Z. WoThis image is classified by VGG16 as cash machine with jna, Rethinking the Inception Architecture for Coma probability of 99%. Looking at the tagged filters, filters puter Vision, in: 2016 IEEE Conference on Comof the tags cash machine are mostly activate, followed puter Vision and Pattern Recognition, CVPR 2016, by screen, CD player, and file . Considering Figure 5 (b) 2016, pp. 2818–2826. and having knowledge about the other images of the [3] G. Montavon, W. Samek, K.-R. Müller, Methods class computer in ImageNet, the reason this image is not for interpreting and understanding deep neural netassigned to this class becomes clear. Generally, frontal works, Digital Signal Processing 73 (2018) 1–15. images of a computer were used for the computer class [4] C. Olah, A. Mordvintsev, L. Schubert, Feature Visufor learning. However, this image does not correspond alization, Distill (2017). to the same distribution. Thus, it is dificult for the

References

[5] C. Olah , A. Satyanarayan , I. Johnson, S. Carter , L. Schubert , K. Ye , A. Mordvintsev , The Building Blocks of Interpretability, Distill ( 2018 ).

[6] F. Hohman , H. Park , C. Robinson , D. H. P. Chau , Summit: Scaling deep learning interpretability by visualizing activation and attribution summarizations , IEEE transactions on visualization and computer graphics 26 ( 2019 ) 1096 - 1106 .

[7] M. D. Zeiler , R. Fergus , Visualizing and Understanding Convolutional Networks , in: Computer Vision - ECCV 2014 , volume 8689 of Lecture Notes in Computer Science, 2014 , pp. 818 - 833 .

[8] R. Fong , A. Vedaldi, Net2Vec: Quantifying and Explaining How Concepts Are Encoded by Filters in Deep Neural Networks , in: Conference on Computer Vision and Pattern Recognition, CVPR 2018 , 2018 , pp. 8730 - 8738 .

[9] D. Bau , B. Zhou , A. Khosla , A. Oliva , A. Torralba , Network Dissection: Quantifying Interpretability of Deep Visual Representations , in: Conference on Computer Vision and Pattern Recognition, CVPR 2017 , 2017 , pp. 3319 - 3327 .

[10] R. B. Girshick , J. Donahue , T. Darrell , J. Malik, Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation , in: Conference on Computer Vision and Pattern Recognition, CVPR 2014 , 2014 , pp. 580 - 587 .

[11] K. Simonyan , A. Vedaldi , A. Zisserman , Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps , in: 2nd International Conference on Learning Representations, ICLR 2014 , 2014 .

[12] J. T. Springenberg , A. Dosovitskiy , T. Brox , M. A. Riedmiller , Striving for Simplicity: The All Convolutional Net , in: 3rd International Conference on Learning Representations, ICLR 2015 , 2015 .

[13] R. Speer , J. Chin , C. Havasi, ConceptNet 5 . 5: An Open Multilingual Graph of General Knowledge , in: Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence , 2017 , pp. 4444 - 4451 .

[14] A. Nguyen , T. Weller , M. Färber , Y. Sure-Vetter , Making Neural Networks FAIR, in: Knowledge Graphs and Semantic Web - Second Iberoamerican Conference and First Indo-American Conference , KGSWC 2020 , volume 1232 of Communications in Computer and Information Science, 2020 , pp. 29 - 44 .

[15] D. Erhan , Y. Bengio , A. Courville , P. Vincent , Visualizing Higher-Layer Features of a Deep Network , Technical Report , Univeristé de Montréal ( 2009 ).

[16] O. Russakovsky , J. Deng , H. Su , J. Krause , S. Satheesh , S. Ma , Z. Huang , A. Karpathy , A. Khosla , M. Bernstein , A. C. Berg , L. Fei-Fei, ImageNet Large Scale Visual Recognition Challenge , International Journal of Computer Vision (IJCV) 115 ( 2015 ) 211 - 252 .