Large Language Models (LLMs) have achieved unprecedented breakthroughs in various natural language processing domains. However, the enigmatic ``black-box'' nature of LLMs remains a significant challenge for interpretability, hampering transparent and accountable applications. While past approaches, such as attention visualization, pivotal subnetwork extraction, and concept-based analyses, offer some insight, they often focus on either local or global explanations within a single dimension, occasionally falling short in providing comprehensive clarity. In response, we propose a novel methodology anchored in sparsity-guided techniques, aiming to provide a holistic interpretation of LLMs. Our framework, termed SparseCBM, innovatively integrates sparsity to elucidate three intertwined layers of interpretation: input, subnetwork, and concept levels. In addition, the newly introduced dimension of interpretable inference-time intervention facilitates dynamic adjustments to the model during deployment. Through rigorous empirical evaluations on real-world datasets, we demonstrate that SparseCBM delivers a profound understanding of LLM behaviors, setting it apart in both interpreting and ameliorating model inaccuracies. Codes are provided in supplements.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Discriminative learning effectively predicts true object class for image classification. However, it often results in false positives for outliers, posing critical concerns in applications like autonomous driving and video surveillance systems. Previous attempts to address this challenge involved training image classifiers through contrastive learning using actual outlier data or synthesizing outliers for self-supervised learning. Furthermore, unsupervised generative modeling of inliers in pixel space has shown limited success for outlier detection. In this work, we introduce a quantile-based maximum likelihood objective for learning the inlier distribution to improve the outlier separation during inference. Our approach fits a normalizing flow to pre-trained discriminative features and detects the outliers according to the evaluated log-likelihood. The experimental evaluation demonstrates the effectiveness of our method as it surpasses the performance of the state-of-the-art unsupervised methods for outlier detection. The results are also competitive compared with a recent self-supervised approach for outlier detection. Our work allows to reduce dependency on well-sampled negative training data, which is especially important for domains like medical diagnostics or remote sensing.
Visual Question Answering (VQA) based on multi-modal data facilitates real-life applications such as home robots and medical diagnoses. One significant challenge is to devise a robust decentralized learning framework for various client models where centralized data collection is refrained due to confidentiality concerns. This work aims to tackle privacy-preserving VQA by decoupling a multi-modal model into representation modules and a contrastive module, leveraging inter-module gradients sharing and inter-client weight sharing. To this end, we propose Bidirectional Contrastive Split Learning (BiCSL) to train a global multi-modal model on the entire data distribution of decentralized clients. We employ the contrastive loss that enables a more efficient self-supervised learning of decentralized modules. Comprehensive experiments are conducted on the VQA-v2 dataset based on five SOTA VQA models, demonstrating the effectiveness of the proposed method. Furthermore, we inspect BiCSL's robustness against a dual-key backdoor attack on VQA. Consequently, BiCSL shows significantly enhanced resilience when exposed to the multi-modal adversarial attack compared to the centralized learning method, which provides a promising approach to decentralized multi-modal learning.
Toward Robustness in Multi-Label Classification: A Data Augmentation Strategy against Imbalance and Noise
PDF ↗Multi-label classification poses challenges due to imbalanced and noisy labels in training data. In this paper, we propose a unified data augmentation method, named BalanceMix, to address these challenges. Our approach includes two samplers for imbalanced labels, generating minority-augmented instances with high diversity. It also refines multi-labels at the label-wise granularity, categorizing noisy labels as clean, re-labeled, or ambiguous for robust optimization. Extensive experiments on three benchmark datasets demonstrate that BalanceMix outperforms existing state-of-the-art methods. We release the code at https://github.com/DISL-Lab/BalanceMix.
There has been much recent interest in developing fair clustering algorithms that seek to do justice to the representation of groups defined along sensitive attributes such as race and sex. Within the centroid clustering paradigm, these algorithms are seen to generate clusterings where different groups are disadvantaged within different clusters with respect to their representativity, i.e., distance to centroid. In view of this deficiency, we propose a novel notion of cluster-level centroid fairness that targets the representativity unfairness borne by groups within each cluster, along with a metric to quantify the same. Towards operationalising this notion, we draw on ideas from political philosophy aligned with consideration for the worst-off group to develop Fair-Centroid; a new clustering method that focusses on enhancing the representativity of the worst-off group within each cluster. Our method uses an iterative optimisation paradigm wherein an initial cluster assignment is refined by reassigning objects to clusters such that the worst-off group in each cluster is benefitted. We compare our notion with a related fairness notion and show through extensive empirical evaluations on real-world datasets that our method significantly enhances cluster-level centroid fairness at low impact on cluster coherence.
Training reinforcement learning (RL) agents to achieve desired goals while also acting morally is a challenging problem. Transformer-based language models (LMs) have shown some promise in moral awareness, but their use in different contexts is problematic because of the complexity and implicitness of human morality. In this paper, we build on text-based games, which are challenging environments for current RL agents, and propose the HuMAL (Human-guided Morality Awareness Learning) algorithm, which adaptively learns personal values through human-agent collaboration with minimal manual feedback. We evaluate HuMAL on the Jiminy Cricket benchmark, a set of text-based games with various scenes and dense morality annotations, using both simulated and actual human feedback. The experimental results demonstrate that with a small amount of human feedback, HuMAL can improve task performance and reduce immoral behavior in a variety of games and is adaptable to different personal values.
Interpretability Benchmark for Evaluating Spatial Misalignment of Prototypical Parts Explanations
PDF ↗Prototypical parts-based networks are becoming increasingly popular due to their faithful self-explanations. However, their similarity maps are calculated in the penultimate network layer. Therefore, the receptive field of the prototype activation region often depends on parts of the image outside this region, which can lead to misleading interpretations. We name this undesired behavior a spatial explanation misalignment and introduce an interpretability benchmark with a set of dedicated metrics for quantifying this phenomenon. In addition, we propose a method for misalignment compensation and apply it to existing state-of-the-art models. We show the expressiveness of our benchmark and the effectiveness of the proposed compensation methodology through extensive empirical studies.
Differentially private (DP) mechanisms have been deployed in a variety of high-impact social settings (perhaps most notably by the U.S. Census). Since all DP mechanisms involve adding noise to results of statistical queries, they are expected to impact our ability to accurately analyze and learn from data, in effect trading off privacy with utility. Alarmingly, the impact of DP on utility can vary significantly among different sub-populations. A simple way to reduce this disparity is with stratification. First compute an independent private estimate for each group in the data set (which may be the intersection of several protected classes), then, to compute estimates of global statistics, appropriately recombine these group estimates. Our main observation is that naive stratification often yields high-accuracy estimates of population-level statistics, without the need for additional privacy budget. We support this observation theoretically and empirically. Our theoretical results center on the private mean estimation problem, while our empirical results center on extensive experiments on private data synthesis to demonstrate the effectiveness of stratification on a variety of private mechanisms. Overall, we argue that this straightforward approach provides a strong baseline against which future work on reducing utility disparities of DP mechanisms should be compared.
Effectively explaining decisions of black-box machine learning models is critical to responsible deployment of AI systems that rely on them. Recognizing their importance, the field of explainable AI (XAI) provides several techniques to generate these explanations. Yet, there is relatively little emphasis on the user (the explainee) in this growing body of work and most XAI techniques generate "one-size-fits-all'' explanations. To bridge this gap and achieve a step closer towards human-centered XAI, we present I-CEE, a framework that provides Image Classification Explanations tailored to User Expertise. Informed by existing work, I-CEE explains the decisions of image classification models by providing the user with an informative subset of training data (i.e., example images), corresponding local explanations, and model decisions. However, unlike prior work, I-CEE models the informativeness of the example images to depend on user expertise, resulting in different examples for different users. We posit that by tailoring the example set to user expertise, I-CEE can better facilitate users' understanding and simulatability of the model. To evaluate our approach, we conduct detailed experiments in both simulation and with human participants (N = 100) on multiple datasets. Experiments with simulated users show that I-CEE improves users' ability to accurately predict the model's decisions (simulatability) compared to baselines, providing promising preliminary results. Experiments with human participants demonstrate that our method significantly improves user simulatability accuracy, highlighting the importance of human-centered XAI.
While modern explanation methods have been shown to be inconsistent and contradictory, the explainability of black-box models nevertheless remains desirable. When the role of explanations extends from understanding models to aiding decision making, the semantics of explanations is not always fully understood – to what extent do explanations ``explain” a decision and to what extent do they merely advocate for a decision? Can we help humans gain insights from explanations accompanying correct predictions and not over-rely on incorrect predictions advocated for by explanations? With this perspective in mind, we introduce the notion of dissenting explanations: conflicting predictions with accompanying explanations. We first explore the advantage of dissenting explanations in the setting of model multiplicity, where multiple models with similar performance may have different predictions. Through a human study on the task of identifying deceptive reviews, we demonstrate that dissenting explanations reduce overreliance on model predictions, without reducing overall accuracy. Motivated by the utility of dissenting explanations we present both global and local methods for their generation.
Warning: this paper contains data, prompts, and model outputs that are offensive in nature. Recently, there has been a surge of interest in integrating vision into Large Language Models (LLMs), exemplified by Visual Language Models (VLMs) such as Flamingo and GPT-4. This paper sheds light on the security and safety implications of this trend. First, we underscore that the continuous and high-dimensional nature of the visual input makes it a weak link against adversarial attacks, representing an expanded attack surface of vision-integrated LLMs. Second, we highlight that the versatility of LLMs also presents visual attackers with a wider array of achievable adversarial objectives, extending the implications of security failures beyond mere misclassification. As an illustration, we present a case study in which we exploit visual adversarial examples to circumvent the safety guardrail of aligned LLMs with integrated vision. Intriguingly, we discover that a single visual adversarial example can universally jailbreak an aligned LLM, compelling it to heed a wide range of harmful instructions (that it otherwise would not) and generate harmful content that transcends the narrow scope of a `few-shot' derogatory corpus initially employed to optimize the adversarial example. Our study underscores the escalating adversarial risks associated with the pursuit of multimodality. Our findings also connect the long-studied adversarial vulnerabilities of neural networks to the nascent field of AI alignment. The presented attack suggests a fundamental adversarial challenge for AI alignment, especially in light of the emerging trend toward multimodality in frontier foundation models.
Counterfactual Explanation (CE) techniques have garnered attention as a means to provide insights to the users engaging with AI systems. While extensively researched in domains such as medical imaging and autonomous vehicles, Graph Counterfactual Explanation (GCE) methods have been comparatively under-explored. GCEs generate a new graph similar to the original one, with a different outcome grounded on the underlying predictive model. Among these GCE techniques, those rooted in generative mechanisms have received relatively limited investigation despite demonstrating impressive accomplishments in other domains, such as artistic styles and natural language modelling. The preference for generative explainers stems from their capacity to generate counterfactual instances during inference, leveraging autonomously acquired perturbations of the input graph. Motivated by the rationales above, our study introduces RSGG-CE, a novel Robust Stochastic Graph Generator for Counterfactual Explanations able to produce counterfactual examples from the learned latent space considering a partially ordered generation sequence. Furthermore, we undertake quantitative and qualitative analyses to compare RSGG-CE's performance against SoA generative explainers, highlighting its increased ability to engendering plausible counterfactual candidates.
Linear temporal logic (LTL) and omega-regular objectives---a superset of LTL---have seen recent use as a way to express non-Markovian objectives in reinforcement learning. We introduce a model-based probably approximately correct (PAC) learning algorithm for omega-regular objectives in Markov decision processes (MDPs). As part of the development of our algorithm, we introduce the epsilon-recurrence time: a measure of the speed at which a policy converges to the satisfaction of the omega-regular objective in the limit. We prove that our algorithm only requires a polynomial number of samples in the relevant parameters, and perform experiments which confirm our theory.
Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training
PDF ↗Traditional adversarial training, while effective at improving machine learning model robustness, is computationally intensive. Fast Adversarial Training (FAT) addresses this by using a single-step attack to generate adversarial examples more efficiently. Nonetheless, FAT is susceptible to a phenomenon known as catastrophic overfitting, wherein the model's adversarial robustness abruptly collapses to zero during the training phase. To address this challenge, recent studies have suggested adopting adversarial initialization with Fast Gradient Sign Method Adversarial Training (FGSM-AT), which recycles adversarial perturbations from prior epochs by computing gradient momentum. However, our research has uncovered a flaw in this approach. Given that data augmentation is employed during the training phase, the samples in each epoch are not identical. Consequently, the method essentially yields not the adversarial perturbation of a singular sample, but rather the Universal Adversarial Perturbation (UAP) of a sample and its data augmentation. This insight has led us to explore the potential of using UAPs for adversarial initialization within the context of FGSM-AT. We have devised various strategies for adversarial initialization utilizing UAPs, including single, class-based, and feature-based UAPs. Experiments conducted on three distinct datasets demonstrate that our method achieves an improved trade-off among robustness, computational cost, and memory footprint. Code is available at https://github.com/fzjcdt/fgsm-uap.
Understanding Likelihood of Normalizing Flow and Image Complexity through the Lens of Out-of-Distribution Detection
PDF ↗Out-of-distribution (OOD) detection is crucial to safety-critical machine learning applications and has been extensively studied. While recent studies have predominantly focused on classifier-based methods, research on deep generative model (DGM)-based methods have lagged relatively. This disparity may be attributed to a perplexing phenomenon: DGMs often assign higher likelihoods to unknown OOD inputs than to their known training data. This paper focuses on explaining the underlying mechanism of this phenomenon. We propose a hypothesis that less complex images concentrate in high-density regions in the latent space, resulting in a higher likelihood assignment in the Normalizing Flow (NF). We experimentally demonstrate its validity for five NF architectures, concluding that their likelihood is untrustworthy. Additionally, we show that this problem can be alleviated by treating image complexity as an independent variable. Finally, we provide evidence of the potential applicability of our hypothesis in another DGM, PixelCNN++.
Explanations in Computer Vision are often desired, but most Deep Neural Networks can only provide saliency maps with questionable faithfulness. Self-Explaining Neural Networks (SENN) extract interpretable concepts with fidelity, diversity, and grounding to combine them linearly for decision-making. While they can explain what was recognized, initial realizations lack accuracy and general applicability. We propose the Quantized-Self-Explaining Neural Network “Q-SENN”. Q-SENN satisfies or exceeds the desiderata of SENN while being applicable to more complex datasets and maintaining most or all of the accuracy of an uninterpretable baseline model, outperforming previous work in all considered metrics. Q-SENN describes the relationship between every class and feature as either positive, negative or neutral instead of an arbitrary number of possible relations, enforcing more binary human-friendly features. Since every class is assigned just 5 interpretable features on average, Q-SENN shows convincing local and global interpretability. Additionally, we propose a feature alignment method, capable of aligning learned features with human language-based concepts without additional supervision. Thus, what is learned can be more easily verbalized. The code is published: https://github.com/ThomasNorr/Q-SENN
Avoiding synthesizing specific visual concepts is an essential challenge in responsible visual synthesis. However, the visual concept that needs to be avoided for responsible visual synthesis tends to be diverse, depending on the region, context, and usage scenarios. In this work, we formalize a new task, Open-vocabulary Responsible Visual Synthesis (ORES), where the synthesis model is able to avoid forbidden visual concepts while allowing users to input any desired content. To address this problem, we present a Two-stage Intervention (TIN) framework. By introducing 1) rewriting with learnable instruction through a large-scale language model (LLM) and 2) synthesizing with prompt intervention on a diffusion synthesis model, it can effectively synthesize images avoiding any concepts but following the user's query as much as possible. To evaluate on ORES, we provide a publicly available dataset, baseline models, and benchmark. Experimental results demonstrate the effectiveness of our method in reducing risks of image generation. Our work highlights the potential of LLMs in responsible visual synthesis. Our code and dataset is public available in https://github.com/kodenii/ORES.
We present a new loss function that addresses the out-of-distribution (OOD) network calibration problem. While many objective functions have been proposed to effectively calibrate models in-distribution, our findings show that they do not always fare well OOD. Based on the Principle of Maximum Entropy, we incorporate helpful statistical constraints observed during training, delivering better model calibration without sacrificing accuracy. We provide theoretical analysis and show empirically that our method works well in practice, achieving state-of-the-art calibration on both synthetic and real-world benchmarks. Our code is available at https://github.com/dexterdley/MaxEnt-Loss.
Current datasets for unwanted social bias auditing are limited to studying protected demographic features such as race and gender. In this work, we introduce a comprehensive benchmark that is meant to capture the amplification of social bias, via stigmas, in generative language models. Taking inspiration from social science research, we start with a documented list of 93 US-centric stigmas and curate a question-answering (QA) dataset which involves simple social situations. Our benchmark, SocialStigmaQA, contains roughly 10K prompts, with a variety of prompt styles, carefully constructed to systematically test for both social bias and model robustness. We present results for SocialStigmaQA with two open source generative language models and we find that the proportion of socially biased output ranges from 45% to 59% across a variety of decoding strategies and prompting styles. We demonstrate that the deliberate design of the templates in our benchmark (e.g., adding biasing text to the prompt or using different verbs that change the answer that indicates bias) impacts the model tendencies to generate socially biased output. Additionally, through manual evaluation, we discover problematic patterns in the generated chain-of-thought output that range from subtle bias to lack of reasoning. Warning: This paper contains examples of text which are toxic, biased, and potentially harmful.
Notions of transition invariants and closure certificates have seen recent use in the formal verification of controlled dynamical systems against \omega-regular properties. Unfortunately, existing approaches face limitations in two directions. First, they require a closed-form mathematical expression representing the model of the system. Such an expression may be difficult to find, too complex to be of any use, or unavailable due to security or privacy constraints. Second, finding such invariants typically rely on optimization techniques such as sum-of-squares (SOS) or satisfiability modulo theory (SMT) solvers. This restricts the classes of systems that need to be formally verified. To address these drawbacks, we introduce a notion of neural closure certificates. We present a data-driven algorithm that trains a neural network to represent a closure certificate. Our approach is formally correct under some mild assumptions, i.e., one is able to formally show that the unknown system satisfies the \omega-regular property of interest if a neural closure certificate can be computed. Finally, we demonstrate the efficacy of our approach with relevant case studies.