Limits...
On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation.

Bach S, Binder A, Montavon G, Klauschen F, Müller KR, Samek W - PLoS ONE (2015)

Bottom Line: Although machine learning methods are solving very successfully a plethora of tasks, they have in most cases the disadvantage of acting as a black box, not providing any information about what made them arrive at a particular decision.These pixel contributions can be visualized as heatmaps and are provided to a human expert who can intuitively not only verify the validity of the classification decision, but also focus further analysis on regions of potential interest.We evaluate our method for classifiers trained on PASCAL VOC 2009 images, synthetic image data containing geometric shapes, the MNIST handwritten digits data set and for the pre-trained ImageNet model available as part of the Caffe open source package.

View Article: PubMed Central - PubMed

Affiliation: Machine Learning Group, Fraunhofer Heinrich Hertz Institute, Berlin, Germany; Machine Learning Group, Technische Universität Berlin, Berlin, Germany.

ABSTRACT
Understanding and interpreting classification decisions of automated image classification systems is of high value in many applications, as it allows to verify the reasoning of the system and provides additional information to the human expert. Although machine learning methods are solving very successfully a plethora of tasks, they have in most cases the disadvantage of acting as a black box, not providing any information about what made them arrive at a particular decision. This work proposes a general solution to the problem of understanding classification decisions by pixel-wise decomposition of nonlinear classifiers. We introduce a methodology that allows to visualize the contributions of single pixels to predictions for kernel-based classifiers over Bag of Words features and for multilayered neural networks. These pixel contributions can be visualized as heatmaps and are provided to a human expert who can intuitively not only verify the validity of the classification decision, but also focus further analysis on regions of potential interest. We evaluate our method for classifiers trained on PASCAL VOC 2009 images, synthetic image data containing geometric shapes, the MNIST handwritten digits data set and for the pre-trained ImageNet model available as part of the Caffe open source package.

No MeSH data available.


Taylor-approximated pixel-wise predictions for a multilayer neural network trained and tested on the MNIST data set.Each group of four horizontally aligned panels shows—from left to right—the input digit, the Taylor root point x0, the gradient of the prediction function f at x0 of a specific digit class indicated by the subscript next to f and the approximated pixel-wise contributions for x.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4498753&req=5

pone.0130140.g011: Taylor-approximated pixel-wise predictions for a multilayer neural network trained and tested on the MNIST data set.Each group of four horizontally aligned panels shows—from left to right—the input digit, the Taylor root point x0, the gradient of the prediction function f at x0 of a specific digit class indicated by the subscript next to f and the approximated pixel-wise contributions for x.

Mentions: Examples for pixel-wise decompositions for the first type of neural networks are given in Figs 11, 12 and 13. Multilayer neural networks were trained on the MNIST [57] data of handwritten digits and solve the posed ten-class problem with a prediction accuracy of 98.25% on the MNIST test set. Our network consists of three linear sum-pooling layers with a bias-inputs, followed by an activation or normalization step each. The first linear layer accepts the 28 × 28 pixel large images as a 784 dimensional input vector and produces a 400-dimensional tanh-activated output vector. The second layer projects those 400 inputs to equally many tanh-activated outputs. The last layer then transforms the 400-dimensional space to a 10-dimensional output space followed by a softmax layer for activation in order to produce output probabilities for each class. The network was trained using a standard error back-propagation algorithm using batches of 25 randomly drawn training samples with an added Gaussian noise layer per training iteration. The above prediction accuracy was achieved after terminating the training procedure after 50 000 iterations.


On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation.

Bach S, Binder A, Montavon G, Klauschen F, Müller KR, Samek W - PLoS ONE (2015)

Taylor-approximated pixel-wise predictions for a multilayer neural network trained and tested on the MNIST data set.Each group of four horizontally aligned panels shows—from left to right—the input digit, the Taylor root point x0, the gradient of the prediction function f at x0 of a specific digit class indicated by the subscript next to f and the approximated pixel-wise contributions for x.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4498753&req=5

pone.0130140.g011: Taylor-approximated pixel-wise predictions for a multilayer neural network trained and tested on the MNIST data set.Each group of four horizontally aligned panels shows—from left to right—the input digit, the Taylor root point x0, the gradient of the prediction function f at x0 of a specific digit class indicated by the subscript next to f and the approximated pixel-wise contributions for x.
Mentions: Examples for pixel-wise decompositions for the first type of neural networks are given in Figs 11, 12 and 13. Multilayer neural networks were trained on the MNIST [57] data of handwritten digits and solve the posed ten-class problem with a prediction accuracy of 98.25% on the MNIST test set. Our network consists of three linear sum-pooling layers with a bias-inputs, followed by an activation or normalization step each. The first linear layer accepts the 28 × 28 pixel large images as a 784 dimensional input vector and produces a 400-dimensional tanh-activated output vector. The second layer projects those 400 inputs to equally many tanh-activated outputs. The last layer then transforms the 400-dimensional space to a 10-dimensional output space followed by a softmax layer for activation in order to produce output probabilities for each class. The network was trained using a standard error back-propagation algorithm using batches of 25 randomly drawn training samples with an added Gaussian noise layer per training iteration. The above prediction accuracy was achieved after terminating the training procedure after 50 000 iterations.

Bottom Line: Although machine learning methods are solving very successfully a plethora of tasks, they have in most cases the disadvantage of acting as a black box, not providing any information about what made them arrive at a particular decision.These pixel contributions can be visualized as heatmaps and are provided to a human expert who can intuitively not only verify the validity of the classification decision, but also focus further analysis on regions of potential interest.We evaluate our method for classifiers trained on PASCAL VOC 2009 images, synthetic image data containing geometric shapes, the MNIST handwritten digits data set and for the pre-trained ImageNet model available as part of the Caffe open source package.

View Article: PubMed Central - PubMed

Affiliation: Machine Learning Group, Fraunhofer Heinrich Hertz Institute, Berlin, Germany; Machine Learning Group, Technische Universität Berlin, Berlin, Germany.

ABSTRACT
Understanding and interpreting classification decisions of automated image classification systems is of high value in many applications, as it allows to verify the reasoning of the system and provides additional information to the human expert. Although machine learning methods are solving very successfully a plethora of tasks, they have in most cases the disadvantage of acting as a black box, not providing any information about what made them arrive at a particular decision. This work proposes a general solution to the problem of understanding classification decisions by pixel-wise decomposition of nonlinear classifiers. We introduce a methodology that allows to visualize the contributions of single pixels to predictions for kernel-based classifiers over Bag of Words features and for multilayered neural networks. These pixel contributions can be visualized as heatmaps and are provided to a human expert who can intuitively not only verify the validity of the classification decision, but also focus further analysis on regions of potential interest. We evaluate our method for classifiers trained on PASCAL VOC 2009 images, synthetic image data containing geometric shapes, the MNIST handwritten digits data set and for the pre-trained ImageNet model available as part of the Caffe open source package.

No MeSH data available.