Researchers are actively trying to gain better insights into the representational properties of convolutional neural networks for guiding better network designs and for interpreting a network's computational nature. Gaining such insights can be an arduous task due to the number of parameters in a network and the complexity of a network's architecture. Current approaches of neural network interpretation include Bayesian probabilistic interpretations and information theoretic interpretations. In this study, we take a different approach to studying convolutional neural networks by proposing an abstract algebraic interpretation using finite transformation semigroup theory. Specifically, convolutional layers are broken up and mapped to a finite space. The state space of the proposed finite transformation semigroup is then defined as a single element within the convolutional layer, with the acting elements defined by surrounding state elements combined with convolution kernel elements. Generators of the finite transformation semigroup are defined to complete the interpretation. We leverage this approach to analyze the basic properties of the resulting finite transformation semigroup to gain insights on the representational properties of convolutional neural networks, including insights into quantized network representation. Such a finite transformation semigroup interpretation can also enable better understanding outside of the confines of fixed lattice data structures, thus useful for handling data that lie on irregular lattices. Furthermore, the proposed abstract algebraic interpretation is shown to be viable for interpreting convolutional operations within a variety of convolutional neural network architectures.
We ask whether the neural network interpretation methods can be fooled via adversarial model manipulation, which is defined as a model fine-tuning step that aims to radically alter the explanations without hurting the accuracy of the original models, e.g., VGG19, ResNet50, and DenseNet121. By incorporating the interpretation results directly in the penalty term of the objective function for fine-tuning, we show that the state-of-the-art saliency map based interpreters, e.g., LRP, Grad-CAM, and SimpleGrad, can be easily fooled with our model manipulation. We propose two types of fooling, Passive and Active, and demonstrate such foolings generalize well to the entire validation set as well as transfer to other interpretation methods. Our results are validated by both visually showing the fooled explanations and reporting quantitative metrics that measure the deviations from the original explanations. We claim that the stability of neural network interpretation method with respect to our adversarial model manipulation is an important criterion to check for developing robust and reliable neural network interpretation method.
Deep neural networks have been shown to be fooled rather easily using adversarial attack algorithms. Practical methods such as adversarial patches have been shown to be extremely effective in causing misclassification. However, these patches are highlighted using standard network interpretation algorithms, thus revealing the identity of the adversary. We show that it is possible to create adversarial patches which not only fool the prediction, but also change what we interpret regarding the cause of the prediction. Moreover, we introduce our attack as a controlled setting to measure the accuracy of interpretation algorithms. We show this using extensive experiments for Grad-CAM interpretation that transfers to occluding patch interpretation as well. We believe our algorithms can facilitate developing more robust network interpretation tools that truly explain the network's underlying decision making process.
A variety of methods have been proposed for interpreting nodes in deep neural networks, which typically involve scoring nodes at lower layers with respect to their effects on the output of higher-layer nodes (where lower and higher layers are closer to the input and output layers, respectively). However, we may be interested in picking out a prioritized collection of subsets of the inputs across a range of scales according to their importance for an output node, and not simply a prioritized ranking across the inputs as singletons. Such a situation may arise in biological applications, for instance, where we are interested in epistatic effects between groups of genes in determining a trait of interest. Here, we outline a flexible framework which may be used to generate multiscale network interpretations, using any previously defined scoring function. We demonstrate the ability of our method to pick out biologically important genes and gene sets in the domains of cancer and psychiatric genomics.
We introduce a novel approach to feed-forward neural network interpretation based on partitioning the space of sequences of neuron activations. In line with this approach, we propose a model-specific interpretation method, called YASENN. Our method inherits many advantages of model-agnostic distillation, such as an ability to focus on the particular input region and to express an explanation in terms of features different from those observed by a neural network. Moreover, examination of distillation error makes the method applicable to the problems with low tolerance to interpretation mistakes. Technically, YASENN distills the network with an ensemble of layer-wise gradient boosting decision trees and encodes the sequences of neuron activations with leaf indices. The finite number of unique codes induces a partitioning of the input space. Each partition may be described in a variety of ways, including examination of an interpretable model (e.g. a logistic regression or a decision tree) trained to discriminate between objects of those partitions. Our experiments provide an intuition behind the method and demonstrate revealed artifacts in neural network decision making.
Network interpretation as an effort to reveal the features learned by a network remains largely visualization-based. In this paper, our goal is to tackle semantic network interpretation at both filter and decision level. For filter-level interpretation, we represent the concepts a filter encodes with a probability distribution of visual attributes. The decision-level interpretation is achieved by textual summarization that generates an explanatory sentence containing clues behind a network's decision. A Bayesian inference algorithm is proposed to automatically associate filters and network decisions with visual attributes. Human study confirms that the semantic interpretation is a beneficial alternative or complement to visualization methods. We demonstrate the crucial role that semantic network interpretation can play in understanding a network's failure patterns. More importantly, semantic network interpretation enables a better understanding of the correlation between a model's performance and its distribution metrics like filter selectivity and concept sparseness.