This paper focuses on the newly emerged research topic, i.e., whether the complex decision-making logic of a DNN can be mathematically summarized into a few simple logics. Beyond the explanation of a static DNN, in this paper, we hope to show that the seemingly complex learning dynamics of a DNN can be faithfully represented as the change of a few primitive interaction patterns encoded by the DNN. Therefore, we redefine the interaction of principal feature components in intermediate-layer features, which enables us to concisely summarize the highly complex dynamics of interactions throughout the learning of the DNN. The mathematical faithfulness of the new interaction is experimentally verified. From the perspective of learning efficiency, we find that the interactions naturally belong to five groups (reliable, withdrawn, forgotten, betraying, and fluctuating interactions), each representing a distinct type of dynamics of an interaction being learned and/or being forgotten. This provides deep insights into the learning process of a DNN.

This paper focuses on the problem of preventing information leakage in neural networks, i.e., assuming that attackers have obtained intermediate-layer features of a neural network, and preventing attackers from inverting these features to the input with private information. We propose a generic method to slightly revise each arbitrary traditional neural network into a multiary-valued rotation-equivariant neural network (RENN) for preventing information leakage. Specifically, we convert real-valued features in the network into multi-ary features, and each element in the feature vector is a multi-ary number. We hide the input information into a certain phase of the multi-ary feature, and rotate the multi-ary feature for attribute obfuscation in the encryption process. The rotation axis and angle can be considered as the private key. In this way, even when attackers have obtained network parameters and intermediate-layer features, they still cannot extract input information without knowing the rotation information. More crucially, the encryption operation does not damage the spatial correlations between features, so that the encrypted features can be easily processed by convolution operations in the neural network without difficulties. In order to implement successful encryption and decryption, the RENN is designed to satisfy the rotation equivariance property. To this end, we propose a set of rules to revise classic operations in the neural network to ensure the rotation equivariance property. Besides, we prove that the $d$d-ary RENN is downward compatible with the $d^{\prime }$d'-ary RENN when $d^{\prime }< d$d'

This paper aims to illustrate the concept-emerging phenomenon in a trained DNN. Specifically, we find that the inference score of a DNN can be disentangled into the effects of a few interactive concepts. These concepts can be understood as causal patterns in a sparse, symbolic causal graph, which explains the DNN. The faithfulness of using such a causal graph to explain the DNN is theoretically guaranteed, because we prove that the causal graph can well mimic the DNN's outputs on an exponential number of different masked samples. Besides, such a causal graph can be further simplified and re-written as an And-Or graph (AOG), without losing much explanation accuracy. The code is released at https://github.com/sjtu-xai-lab/aog.

This paper aims to theoretically analyze the complexity of feature transformations encoded in piecewise linear DNNs with ReLU layers. We propose metrics to measure three types of complexities of transformations based on the information theory. We further discover and prove the strong correlation between the complexity and the disentanglement of transformations. Based on the proposed metrics, we analyze two typical phenomena of the change of the transformation complexity during the training process, and explore the ceiling of a DNN’s complexity. The proposed metrics can also be used as a loss to learn a DNN with the minimum complexity, which also controls the over-fitting level of the DNN and influences adversarial robustness, adversarial transferability, and knowledge consistency. Comprehensive comparative studies have provided new perspectives to understand the DNN. The code is released at https://github.com/sjtu-XAI-lab/transformation-complexity.

This paper aims to define, visualize, and analyze the feature complexity that is learned by a DNN. We propose a generic definition for the feature complexity. Given the feature of a certain layer in the DNN, our method decomposes and visualizes feature components of different complexity orders from the feature. The feature decomposition enables us to evaluate the reliability, the effectiveness, and the significance of over-fitting of these feature components. Furthermore, such analysis helps to improve the performance of DNNs. As a generic method, the feature complexity also provides new insights into existing deep-learning techniques, such as network compression and knowledge distillation.

This paper provides a unified view to explain different adversarial attacks and defense methods, i.e. the view of multi-order interactions between input variables of DNNs. Based on the multi-order interaction, we discover that adversarial attacks mainly affect high-order interactions to fool the DNN. Furthermore, we find that the robustness of adversarially trained DNNs comes from category-specific low-order interactions. Our findings provide a potential method to unify adversarial perturbations and robustness, which can explain the existing robustness-boosting methods in a principle way. Besides, our findings also make a revision of previous inaccurate understanding of the shape bias of adversarially learned features. Our code is available online at https://github.com/Jie-Ren/A-Unified-Game-Theoretic-Interpretation-of-Adversarial-Robustness.

In this paper, we present a method to mine object-part patterns from conv-layers of a pre-trained convolutional neural network (CNN). The mined object-part patterns are organized by an And-Or graph (AOG). This interpretable AOG representation consists of a four-layer semantic hierarchy, i.e., semantic parts, part templates, latent patterns, and neural units. The AOG associates each object part with certain neural units in feature maps of conv-layers. The AOG is constructed with very few annotations (e.g., 3-20) of object parts. We develop a question-answering (QA) method that uses active human-computer communications to mine patterns from a pre-trained CNN, in order to explain features in conv-layers incrementally. During the learning process, our QA method uses the current AOG for part localization. The QA method actively identifies objects, whose feature maps cannot be explained by the AOG. Then, our method asks people to annotate parts on the unexplained objects, and uses answers to discover CNN patterns corresponding to newly labeled parts. In this way, our method gradually grows new branches and refines existing branches on the AOG to semanticize CNN representations. In experiments, our method exhibited a high learning efficiency. Our method used about 1/6- 1/3 of the part annotations for training, but achieved similar or better part-localization performance than fast-RCNN methods.

This paper presents a method to pursue a semantic and quantitative explanation for the knowledge encoded in a convolutional neural network (CNN). The estimation of the specific rationale of each prediction made by the CNN presents a key issue of understanding neural networks, and it is of significant values in real applications. In this study, we propose to distill knowledge from the CNN into an explainable additive model, which explains the CNN prediction quantitatively. We discuss the problem of the biased interpretation of CNN predictions. To overcome the biased interpretation, we develop prior losses to guide the learning of the explainable additive model. Experimental results have demonstrated the effectiveness of our method.