Interpretability Of Deep Neural Networks Research Articles

Interpretability of deep neural networks: A review of methods, classification and hardware

Artificial intelligence, and especially deep neural networks, have evolved substantially in the recent years, infiltrating numerous domains of applications, often greatly impactful to society’s well-being. As a result, the need to understand how these models operate in depth and to access explanations of their decisions has become more vital than ever. Tending to this demand, the following paper aims to provide a thorough overview of the methods that have so far been developed to explain deep neural networks. Key aspects of explainability are defined and a straightforward classification of existing approaches is introduced, along with numerous examples. The task of realizing these methods on hardware is also discussed to complete the understanding of their application.

B-Cos Alignment for Inherently Interpretable CNNs and Vision Transformers.

We present a new direction for increasing the interpretability of deep neural networks (DNNs) by promoting weight-input alignment during training. For this, we propose to replace the linear transformations in DNNs by our novel B-cos transformation. As we show, a sequence (network) of such transformations induces a single linear transformation that faithfully summarises the full model computations. Moreover, the B-cos transformation is designed such that the weights align with relevant signals during optimisation. As a result, those induced linear transformations become highly interpretable and highlight task-relevant features. Importantly, the B-cos transformation is designed to be compatible with existing architectures and we show that it can easily be integrated into virtually all of the latest state of the art models for computer vision-e.g. ResNets, DenseNets, ConvNext models, as well as Vision Transformers-by combining the B-cos-based explanations with normalisation and attention layers, all whilst maintaining similar accuracy on ImageNet. Finally, we show that the resulting explanations are of high visual quality and perform well under quantitative interpretability metrics.

Local Interpretations for Explainable Natural Language Processing: A Survey

As the use of deep learning techniques has grown across various fields over the past decade, complaints about the opaqueness of the black-box models have increased, resulting in an increased focus on transparency in deep learning models. This work investigates various methods to improve the interpretability of deep neural networks for Natural Language Processing (NLP) tasks, including machine translation and sentiment analysis. We provide a comprehensive discussion on the definition of the term interpretability and its various aspects at the beginning of this work. The methods collected and summarised in this survey are only associated with local interpretation and are specifically divided into three categories: (1) interpreting the model’s predictions through related input features; (2) interpreting through natural language explanation; (3) probing the hidden states of models and word representations.

Augmented Grad-CAM++: Super-Resolution Saliency Maps for Visual Interpretation of Deep Neural Network

In recent years, deep neural networks have shown superior performance in various fields, but interpretability has always been the Achilles’ heel of deep neural networks. The existing visual interpretation methods for deep neural networks still suffer from inaccurate and insufficient target localization and low-resolution saliency maps. To address the above issues, this paper presents a saliency map generation method based on image geometry augmentation and super-resolution called augmented high-order gradient weighting class activation mapping (augmented grad-CAM++). Unlike previous approaches that rely on a single input image to generate saliency maps, this method first introduces the image geometry augmentation technique to create a set of augmented images for the input image and generate activation mappings separately. Secondly, the augmented activation mappings are combined to form the final saliency map. Finally, a super-resolution technique is introduced to add pixel points to reconstruct the saliency map pixels to improve the resolution of the saliency map. The proposed method is applied to analyze standard image data and industrial surface defect images. The results indicate that, in experiments conducted on standard image data, the proposed method achieved a 3.1% improvement in the accuracy of capturing target objects compared to traditional methods. Furthermore, the resolution of saliency maps was three times higher than that of traditional methods. In the application of industrial surface defect detection, the proposed method demonstrated an 11.6% enhancement in the accuracy of capturing target objects, concurrently reducing the false positive rate. The presented approach enables more accurate and comprehensive capture of target objects with higher resolution, thereby enhancing the visual interpretability of deep neural networks. This improvement contributes to the greater interpretability of deep learning models in industrial applications, offering substantial performance gains for the practical deployment of deep learning networks in the industrial domain.

APGVAE: Adaptive disentangled representation learning with the graph-based structure information

Neural networks are used to learn task-oriented high-level representations in an end-to-end manner by building a multi-layer neural network. Generation models have developed rapidly with the emergence of deep neural networks. But it still has problems with the insufficient authenticity of generated images, the deficiency of diversity, consistency, and unexplainability in the generation process. Disentangled representation is an effective method to learn a high-level feature representation and realize the interpretability of deep neural networks. We propose a general disentangled representation learning network with variational autoencoder network as the basic framework for the image generation process. The graph-based structure of the priors is embedded in the last module of the deep encoder network to build the feature spaces by the class, task-oriented, and task-unrelated information respectively. Meanwhile the priors should be adaptively modified with the task relevance of a generated image. And the semi-supervised learning is further involved in the disentangled representation network framework to reduce the requirements of label and extend the majority of feature space under the task-unrelated feature assumption. Experimental results show that the proposed method is efficient for various types of images and has a good potential for further research and development.

SHAMSUL: Systematic Holistic Analysis to investigate Medical Significance Utilizing Local interpretability methods in deep learning for chest radiography pathology prediction

The interpretability of deep neural networks has become a subject of great interest within the medical and healthcare domain. This attention stems from concerns regarding transparency, legal and ethical considerations, and the medical significance of predictions generated by these deep neural networks in clinical decision support systems. To address this matter, our study delves into the application of four well-established interpretability methods: Local Interpretable Model-agnostic Explanations (LIME), Shapley Additive exPlanations (SHAP), Gradient-weighted Class Activation Mapping (Grad-CAM), and Layer-wise Relevance Propagation (LRP). Leveraging the approach of transfer learning with a multi-label-multi-class chest radiography dataset, we aim to interpret predictions pertaining to specific pathology classes. Our analysis encompasses both single-label and multi-label predictions, providing a comprehensive and unbiased assessment through quantitative and qualitative investigations, which are compared against human expert annotation. Notably, Grad-CAM demonstrates the most favorable performance in quantitative evaluation, while the LIME heatmap score segmentation visualization exhibits the highest level of medical significance. Our research underscores both the outcomes and the challenges faced in the holistic approach adopted for assessing these interpretability methods and suggests that a multimodal-based approach, incorporating diverse sources of information beyond chest radiography images, could offer additional insights for enhancing interpretability in the medical domain.

A Survey of Controllable Text Generation Using Transformer-based Pre-trained Language Models

Controllable Text Generation (CTG) is an emerging area in the field of natural language generation (NLG). It is regarded as crucial for the development of advanced text generation technologies that better meet the specific constraints in practical applications. In recent years, methods using large-scale pre-trained language models (PLMs), in particular the widely used Transformer-based PLMs, have become a new paradigm of NLG, allowing generation of more diverse and fluent text. However, due to the limited level of interpretability of deep neural networks, the controllability of these methods needs to be guaranteed. To this end, controllable text generation using Transformer-based PLMs has become a rapidly growing yet challenging new research hotspot. A diverse range of approaches have emerged in the past 3 to 4 years, targeting different CTG tasks that require different types of controlled constraints. In this article, we present a systematic critical review on the common tasks, main approaches, and evaluation methods in this area. Finally, we discuss the challenges that the field is facing, and put forward various promising future directions. To the best of our knowledge, this is the first survey article to summarize the state-of-the-art CTG techniques from the perspective of Transformer-based PLMs. We hope it can help researchers and practitioners in the related fields to quickly track the academic and technological frontier, providing them with a landscape of the area and a roadmap for future research.

Jacobian norm with Selective Input Gradient Regularization for interpretable adversarial defense

Deep neural networks (DNNs) can be easily deceived by imperceptible alterations known as adversarial examples. These examples can lead to misclassification, posing a significant threat to the reliability of deep learning systems in real-world applications. Adversarial training (AT) is a popular technique used to enhance robustness by training models on a combination of corrupted and clean data. However, existing AT-based methods often struggle to handle transferred adversarial examples that can fool multiple defense models, thereby falling short of meeting the generalization requirements for real-world scenarios. Furthermore, AT typically fails to provide interpretable predictions, which are crucial for domain experts seeking to understand the behavior of DNNs. To overcome these challenges, we present a novel approach called Jacobian norm and Selective Input Gradient Regularization (J-SIGR). Our method leverages Jacobian normalization to improve robustness and introduces regularization of perturbation-based saliency maps, enabling interpretable predictions. By adopting J-SIGR, we achieve enhanced defense capabilities and promote high interpretability of DNNs. We evaluate the effectiveness of J-SIGR across various architectures by subjecting it to powerful adversarial attacks. Our experimental evaluations provide compelling evidence of the efficacy of J-SIGR against transferred adversarial attacks, while preserving interpretability. The project code can be found at https://github.com/Lywu-github/jJ-SIGR.git.

A graph-based interpretability method for deep neural networks

With the development of artificial intelligence, the most representative deep learning has been applied to various fields, which is greatly influencing human society. However, deep neural networks (DNNs) are still a black-box model, and the process how they make decisions internally is still difficult to understand and control. At the same time, DNNs take up more hardware resources, resulting in high energy consumption. Therefore, it is significant to study the characteristics of deep AI models and deeply understand the interactions between parameters within AI models so as to improve the interpretability of DNNs, optimize their structure and increase their computational efficiency. In this paper, we propose a graph-based interpretability method for deep neural networks (GIMDNN). The running parameters of DNNs are modeled as a graph by using a kernel function or the Graph Transformer Networks (GTN), where the nodes of the graph are obtained by dimensional mapping of the parameters of the DNNs, and the edges are calculated by the Gaussian kernel function. The generated graphs are classified by a graph convolutional network (GCN). The association relationship between the adjacent layers and the running mechanism of DNNs are analyzed, and the importance of the parameters of each layer in the DNNs for the final classification result can be obtained. Convolutional neural networks (CNNs) are one of the most representative models in DNNs. The proposed method is experimentally evaluated on the CNNs. The experimental results show that the proposed method can interpret the associations among the weight parameters as well as the correlation between two adjacent layers. Therefore, the DNNs for special tasks, such as portable applications, edge computing, and so on, can be customized, the number of parameters can be reduced. It is valuable to interpret the operation and principle of CNNs.

Interpretability for reliable, efficient, and self-cognitive DNNs: From theories to applications

In recent years, remarkable achievements have been made in artificial intelligence tasks and applications based on deep neural networks (DNNs), especially in the fields of vision, speech, text, and multimodal analysis. The learning of DNNs is not only the process of abstracting essential laws from data but also the result of nonlinear fitting from massive high-dimensional data. However, the architecture, operation mode, and learning ability of DNNs are still far from those of human brain neurons, and the calculation and reasoning are extremely complex, making the model’s analysis and interpretation crucial. To free DNNs from their dependence on complex structures and massive data, a lot of related works toward the interpretability of DNNs have been proposed. In this review, we elaborate on the definition of model interpretability from the three perspectives of model reliability, feature efficiency, and self-cognition. The interpretability theory of DNNs is summarized from four aspects: model adversarial attack and defense, feature representations, information and geometry, and causal counterfactual. In addition, we categorize the interpretable methods involved according to typical application scenarios. Finally, we discuss the research goals that have not yet been achieved. We sincerely hope that our work will benefit the field and attract more researchers to devote their energy to the interpretability of DNNs, thereby pushing forward the long-term development of artificial neural networks and artificial intelligence.

Clean-label poisoning attack with perturbation causing dominant features

Numerous studies have proved that deep learning models are sensitive to attacks from adversarial examples. However, due to the high computationally demanding and limiting explanation for deep neural networks (DNNs), research on poisoning attacks against deep models is insufficient. In this work, considering the interpretation that DNNs are feature extractors, we propose a clean-label poisoning attack against DNNs by dominant image features added to the original data. Each category of clean data is augmented with imperceptible perturbations which we call “dominant image features” of another category. The dominant image features determine which class the input belongs to, while the clean input behaves like noise. Well-trained DNNs on these poisoning data will get extremely low accuracy on clean test data. The attack can make personal data unlearnable for DNNs, which preserves an individual's privacy. Besides, our data poisoning attack can be expanded into a powerful clean-label backdoor attack in that the poisoned samples and their labels are consistent for humans. Experimental results show that our proposed attack can exceed the previous advanced poisoning attacks which demonstrate the effectiveness of our presented attack. Additionally, our work provides a new perspective on the interpretation of DNNs as feature extractors.

Eye tracking based deep learning analysis for the early detection of diabetic retinopathy: A pilot study

Deep neural networks (DNNs) have exhibited impressive performance in the diabetic retinopathy (DR) computer-aided diagnosis (CAD) systems. However, DNNs are quite hungry for enormous labeled images. The limited data volume may degrade both the accuracy and interpretability of DNNs seriously. To alleviate this situation, it is significant to excavate valuable prior information. We aim to explore how to utilize ophthalmologist’s eye tracking information into an early DR detection system thus to improve the classification accuracy and interpretability. In this paper, ophthalmologists’ gaze maps are firstly collected from their eye movements through eye tracker during DR diagnosis. Then we investigate an eye tracking based early DR detection model based on ophthalmologists’ gaze maps. First, we analysis the effect of the gaze map integrated with the original fundus image based on two image fusion approaches. Second, the weighted gaze map is regarded as a supervised mask to guide the learning of the attention of a DNN model. Additionally, we propose a novel difficulty-aware and class-adaptive gaze map attention learning strategy to enhance the model interpretability. Comparative experiments prove that the weighted gaze map contains more medical knowledge for diagnostic decision. Meanwhile, the attention guidance method via class activate map (CAM) regularization demonstrates its superiority on improving both the accuracy and interpretability of early DR detection model. These investigations indicate that ophthalmologists’ gaze maps, as medical prior knowledge, can contribute to the design of early DR detection model.

Deep learning for studying drawing behavior: A review.

In recent years, computer science has made major advances in understanding drawing behavior. Artificial intelligence, and more precisely deep learning, has displayed unprecedented performance in the automatic recognition and classification of large databases of sketches and drawings collected through touchpad devices. Although deep learning can perform these tasks with high accuracy, the way they are performed by the algorithms remains largely unexplored. Improving the interpretability of deep neural networks is a very active research area, with promising recent advances in understanding human cognition. Deep learning thus offers a powerful framework to study drawing behavior and the underlying cognitive processes, particularly in children and non-human animals, on whom knowledge is incomplete. In this literature review, we first explore the history of deep learning as applied to the study of drawing along with the main discoveries in this area, while proposing open challenges. Second, multiple ideas are discussed to understand the inherent structure of deep learning models. A non-exhaustive list of drawing datasets relevant to deep learning approaches is further provided. Finally, the potential benefits of coupling deep learning with comparative cultural analyses are discussed.

Achieving efficient interpretability of reinforcement learning via policy distillation and selective input gradient regularization

Although deep Reinforcement Learning (RL) has proven successful in a wide range of tasks, one challenge it faces is interpretability when applied to real-world problems. Saliency maps are frequently used to provide interpretability for deep neural networks. However, in the RL domain, existing saliency map approaches are either computationally expensive and thus cannot satisfy the real-time requirement of real-world scenarios or cannot produce interpretable saliency maps for RL policies. In this work, we propose an approach of Distillation with selective Input Gradient Regularization (DIGR) which uses policy distillation and input gradient regularization to produce new policies that achieve both high interpretability and computation efficiency in generating saliency maps. Our approach is also found to improve the robustness of RL policies to multiple adversarial attacks. We conduct experiments on three tasks, MiniGrid (Fetch Object), Atari (Breakout) and CARLA Autonomous Driving, to demonstrate the importance and effectiveness of our approach.

Deep Neural Network Denoising Model Based on Sparse Representation Algorithm for ECG Signal

Electrocardiogram (ECG) denoising is very important for heart diseases diagnosis. The traditional ECG denoising models have problems such as single noise type and poor interpretability of deep neural networks. The innovation of the proposed method is to incorporate the precious achievements of traditional methods into the design of neural networks and to build a bridge between them. Therefore, a novel interpretable deep denoising framework based on sparse representation is proposed in this study, and the half quadratic splitting (HQS) algorithm is applied to decompose the denoising method into sparse representations as an iterative solution process. In addition, a new weight distribution module is designed to extract adpative hyperparameters based on ECG correlation instead of empirical values and greatly improves the efficiency of hyperparameter selection. To demonstrate the fairness and the effectiveness of the proposed method, four different denoising models with different data preprocessing technique are used for comparison. The extensive experimental validation and simulation studies demonstrated that the proposed framework has excellent performance in quantitative and visual evaluation.

Functional network: A novel framework for interpretability of deep neural networks

The layered structure of deep neural networks hinders the use of numerous analysis tools and thus the development of its interpretability. Inspired by the success of functional brain networks, we propose a novel framework for interpretability of deep neural networks, that is, the functional network. We construct the functional network of fully connected networks and explore its small-worldness. In our experiments, the mechanisms of regularization methods, namely, batch normalization and dropout, are revealed using graph theoretical analysis and topological data analysis. Our empirical analysis shows the following: (1) batch normalization enhances model performance by increasing the global efficiency and the number of loops but reduces adversarial robustness by lowering the fault tolerance. (2) Dropout improves generalization and robustness of models by improving the functional specialization and fault tolerance. (3) The models with different regularizations can be clustered accurately according to their functional topological differences, reflecting the great potential of the functional network and topological data analysis in interpretability.

Semantic Interpretation for Convolutional Neural Networks: What Makes a Cat a Cat?

The interpretability of deep neural networks has attracted increasing attention in recent years, and several methods have been created to interpret the "black box" model. Fundamental limitations remain, however, that impede the pace of understanding the networks, especially the extraction of understandable semantic space. In this work, the framework of semantic explainable artificial intelligence(S-XAI) is introduced, which utilizes a sample compression method based on the distinctive row-centered principal component analysis (PCA) that is different from the conventional column-centered PCA to obtain common traits of samples from the convolutional neural network (CNN), and extracts understandable semantic spaces on the basis of discovered semantically sensitive neurons and visualization techniques. Statistical interpretation of the semantic space is also provided, and the concept of semantic probability is proposed. The experimental results demonstrate that S-XAI is effective in providing a semantic interpretation for the CNN, and offers broad usage, including trustworthiness assessment and semantic sample searching.

Federated Neural Architecture Search for Medical Data Security

Medical data widely exist in the hospital and personal life, usually across institutions and regions. They have essential diagnostic value and therapeutic significance. The disclosure of patient information causes people’s panic, therefore, medical data security solution is very crucial for intelligent health care. The emergence of federated learning (FL) provides an effective solution, which only transmits model parameters, breaking through the bottleneck of medical data sharing, protecting data security, and avoiding economic losses. Meanwhile, the neural architecture search (NAS) has become a popular method to automatically search the optimal neural architecture for solving complex practical problems. However, few papers have combined the FL and NAS for simultaneous privacy protection and model architecture selection. Convolutional neural network (CNN) has outstanding performance in the image recognition field. Combining CNN and fuzzy rough sets can effectively improve the interpretability of deep neural networks. This article aims to develop a multiobjective convolutional interval type-2 fuzzy rough FL model based on NAS (CIT2FR-FL-NAS) for medical data security with an improved multiobjective evolutionary algorithm. We test the proposed framework on the LC25000 lung and colon histopathological image dataset. Experimental verification demonstrates that the designed multiobjective CIT2FR-FL-NAS framework can achieve high accuracy superior to state-of-the-art models and reduce network complexity under the condition of protecting medical data security.

Interpretability-Guided Defense Against Backdoor Attacks to Deep Neural Networks

As an emerging threat to deep neural networks (DNNs), backdoor attacks have received increasing attentions due to the challenges posed by the lack of transparency inherent in DNNs. In this article, we develop an efficient algorithm from the interpretability of DNNs to defend against backdoor attacks to DNN models. To extract critical neurons, we deploy sets of control gates following neurons in layers, and the function of a DNN model can be interpreted as semantic sensitivities of neurons to input samples. A backdoor identification approach, derived from the activation frequency distribution on critical neurons, is proposed to reveal anomalies of particular neurons produced by backdoor attacks. Subsequently, a feasible and fine-grained pruning strategy is introduced to eliminate backdoors hidden in DNN models, without the need of retraining. Extensive experiments demonstrate that the proposed algorithm can identify and eliminate malicious backdoors efficiently in both single-target and multitarget scenarios with the performance of a DNN model retained to a large extent.

Microbiome-based disease prediction with multimodal variational information bottlenecks.

Scientific research is shedding light on the interaction of the gut microbiome with the human host and on its role in human health. Existing machine learning methods have shown great potential in discriminating healthy from diseased microbiome states. Most of them leverage shotgun metagenomic sequencing to extract gut microbial species-relative abundances or strain-level markers. Each of these gut microbial profiling modalities showed diagnostic potential when tested separately; however, no existing approach combines them in a single predictive framework. Here, we propose the Multimodal Variational Information Bottleneck (MVIB), a novel deep learning model capable of learning a joint representation of multiple heterogeneous data modalities. MVIB achieves competitive classification performance while being faster than existing methods. Additionally, MVIB offers interpretable results. Our model adopts an information theoretic interpretation of deep neural networks and computes a joint stochastic encoding of different input data modalities. We use MVIB to predict whether human hosts are affected by a certain disease by jointly analysing gut microbial species-relative abundances and strain-level markers. MVIB is evaluated on human gut metagenomic samples from 11 publicly available disease cohorts covering 6 different diseases. We achieve high performance (0.80 < ROC AUC < 0.95) on 5 cohorts and at least medium performance on the remaining ones. We adopt a saliency technique to interpret the output of MVIB and identify the most relevant microbial species and strain-level markers to the model’s predictions. We also perform cross-study generalisation experiments, where we train and test MVIB on different cohorts of the same disease, and overall we achieve comparable results to the baseline approach, i.e. the Random Forest. Further, we evaluate our model by adding metabolomic data derived from mass spectrometry as a third input modality. Our method is scalable with respect to input data modalities and has an average training time of < 1.4 seconds. The source code and the datasets used in this work are publicly available.