Mixture-of-Experts (MoE) models are designed to enhance the efficiency of large language models (LLMs) without proportionally increasing the computational demands. However, their deployment on edge devices still faces significant challenges due to high on-demand loading overheads from managing sparsely activated experts. This paper introduces AdapMoE, an algorithm-system co-design framework for efficient MoE inference. AdapMoE features adaptive expert gating and management to reduce the on-demand loading overheads. We observe the heterogeneity of experts loading across layers and tokens, based on which we propose a sensitivity-based strategy to adjust the number of activated experts dynamically. Meanwhile, we also integrate advanced prefetching and cache management techniques to further reduce the loading latency. Through comprehensive evaluations on various platforms, we demonstrate AdapMoE consistently outperforms existing techniques, reducing the average number of activated experts by 25% and achieving a 1.35x speedup without accuracy degradation. Code is available at: https://github.com/PKU-SEC-Lab/AdapMoE.
Run to run variability in parallel programs caused by floating-point non-associativity has been known to significantly affect reproducibility in iterative algorithms, due to accumulating errors. Non-reproducibility can critically affect the efficiency and effectiveness of correctness testing for stochastic programs. Recently, the sensitivity of deep learning training and inference pipelines to floating-point non-associativity has been found to sometimes be extreme. It can prevent certification for commercial applications, accurate assessment of robustness and sensitivity, and bug detection. New approaches in scientific computing applications have coupled deep learning models with high-performance computing, leading to an aggravation of debugging and testing challenges. Here we perform an investigation of the statistical properties of floating-point non-associativity within modern parallel programming models, and analyze performance and productivity impacts of replacing atomic operations with deterministic alternatives on GPUs. We examine the recently-added deterministic options in PyTorch within the context of GPU deployment for deep learning, uncovering and quantifying the impacts of input parameters triggering run to run variability and reporting on the reliability and completeness of the documentation. Finally, we evaluate the strategy of exploiting automatic determinism that could be provided by deterministic hardware, using the Groq accelerator for inference portions of the deep learning pipeline. We demonstrate the benefits that a hardware-based strategy can provide within reproducibility and correctness efforts.
Stochastic multi-agent multi-armed bandits typically assume that the rewards from each arm follow a fixed distribution, regardless of which agent pulls the arm. However, in many real-world settings, rewards can depend on the sensitivity of each agent to their environment. In medical screening, disease detection rates can vary by test type; in preference matching, rewards can depend on user preferences; and in environmental sensing, observation quality can vary across sensors. Since past work does not specify how to allocate agents of heterogeneous but known sensitivity of these types in a stochastic bandit setting, we introduce a UCB-style algorithm, Min-Width, which aggregates information from diverse agents. In doing so, we address the joint challenges of (i) aggregating the rewards, which follow different distributions for each agent-arm pair, and (ii) coordinating the assignments of agents to arms. Min-Width facilitates efficient collaboration among heterogeneous agents, exploiting the known structure in the agents' reward functions to weight their rewards accordingly. We analyze the regret of Min-Width and conduct pseudo-synthetic and fully synthetic experiments to study the performance of different levels of information sharing. Our results confirm that the gains to modeling agent heterogeneity tend to be greater when the sensitivities are more varied across agents, while combining more information does not always improve performance.
The semantic similarity between documents of a text corpus can be visualized using map-like metaphors based on two-dimensional scatterplot layouts. These layouts result from a dimensionality reduction on the document-term matrix or a representation within a latent embedding, including topic models. Thereby, the resulting layout depends on the input data and hyperparameters of the dimensionality reduction and is therefore affected by changes in them. Furthermore, the resulting layout is affected by changes in the input data and hyperparameters of the dimensionality reduction. However, such changes to the layout require additional cognitive efforts from the user. In this work, we present a sensitivity study that analyzes the stability of these layouts concerning (1) changes in the text corpora, (2) changes in the hyperparameter, and (3) randomness in the initialization. Our approach has two stages: data measurement and data analysis. First, we derived layouts for the combination of three text corpora and six text embeddings and a grid-search-inspired hyperparameter selection of the dimensionality reductions. Afterward, we quantified the similarity of the layouts through ten metrics, concerning local and global structures and class separation. Second, we analyzed the resulting 42817 tabular data points in a descriptive statistical analysis. From this, we derived guidelines for informed decisions on the layout algorithm and highlight specific hyperparameter settings. We provide our implementation as a Git repository at https://github.com/hpicgs/Topic-Models-and-Dimensionality-Reduction-Sensitivity-Study and results as Zenodo archive at https://doi.org/10.5281/zenodo.12772898.
The ubiquity of missing data has sparked considerable attention and focus on tabular data imputation methods. Diffusion models, recognized as the cutting-edge technique for data generation, demonstrate significant potential in tabular data imputation tasks. However, in pursuit of diversity, vanilla diffusion models often exhibit sensitivity to initialized noises, which hinders the models from generating stable and accurate imputation results. Additionally, the sparsity inherent in tabular data poses challenges for diffusion models in accurately modeling the data manifold, impacting the robustness of these models for data imputation. To tackle these challenges, this paper introduces an advanced diffusion model named Self-supervised imputation Diffusion Model (SimpDM for brevity), specifically tailored for tabular data imputation tasks. To mitigate sensitivity to noise, we introduce a self-supervised alignment mechanism that aims to regularize the model, ensuring consistent and stable imputation predictions. Furthermore, we introduce a carefully devised state-dependent data augmentation strategy within SimpDM, enhancing the robustness of the diffusion model when dealing with limited data. Extensive experiments demonstrate that SimpDM matches or outperforms state-of-the-art imputation methods across various scenarios.
Diagnosing and treating abnormalities in the wrist, specifically distal radius, and ulna fractures, is a crucial concern among children, adolescents, and young adults, with a higher incidence rate during puberty. However, the scarcity of radiologists and the lack of specialized training among medical professionals pose a significant risk to patient care. This problem is further exacerbated by the rising number of imaging studies and limited access to specialist reporting in certain regions. This highlights the need for innovative solutions to improve the diagnosis and treatment of wrist abnormalities. Automated wrist fracture detection using object detection has shown potential, but current studies mainly use two-stage detection methods with limited evidence for single-stage effectiveness. This study employs state-of-the-art single-stage deep neural network-based detection models YOLOv5, YOLOv6, YOLOv7, and YOLOv8 to detect wrist abnormalities. Through extensive experimentation, we found that these YOLO models outperform the commonly used two-stage detection algorithm, Faster R-CNN, in fracture detection. Additionally, compound-scaled variants of each YOLO model were compared, with YOLOv8m demonstrating a highest fracture detection sensitivity of 0.92 and mean average precision (mAP) of 0.95. On the other hand, YOLOv6m achieved the highest sensitivity across all classes at 0.83. Meanwhile, YOLOv8x recorded the highest mAP of 0.77 for all classes on the GRAZPEDWRI-DX pediatric wrist dataset, highlighting the potential of single-stage models for enhancing pediatric wrist imaging.
We present many-body Message Passing Neural Network (MPNN) framework that models higher-order node interactions ($\ge 2$ nodes). We model higher-order terms as tree-shaped motifs, comprising a central node with its neighborhood, and apply localized spectral filters on motif Laplacian, weighted by global edge Ricci curvatures. We prove our formulation is invariant to neighbor node permutation, derive its sensitivity bound, and bound the range of learned graph potential. We run regression on graph energies to demonstrate that it scales well with deeper and wider network topology, and run classification on synthetic graph datasets with heterophily and show its consistently high Dirichlet energy growth. We open-source our code at https://github.com/JThh/Many-Body-MPNN.
Development of artificial intelligence (AI) for medical imaging demands curation and cleaning of large-scale clinical datasets comprising hundreds of thousands of images. Some modalities, such as mammography, contain highly standardized imaging. In contrast, breast ultrasound imaging (BUS) can contain many irregularities not indicated by scan metadata, such as enhanced scan modes, sonographer annotations, or additional views. We present an open-source software solution for automatically processing clinical BUS datasets. The algorithm performs BUS scan filtering (flagging of invalid and non-B-mode scans), cleaning (dual-view scan detection, scan area cropping, and caliper detection), and knowledge extraction (BI-RADS Labeling and Measurement fields) from sonographer annotations. Its modular design enables users to adapt it to new settings. Experiments on an internal testing dataset of 430 clinical BUS images achieve >95% sensitivity and >98% specificity in detecting every type of text annotation, >98% sensitivity and specificity in detecting scans with blood flow highlighting, alternative scan modes, or invalid scans. A case study on a completely external, public dataset of BUS scans found that BUSClean identified text annotations and scans with blood flow highlighting with 88.6% and 90.9% sensitivity and 98.3% and 99.9% specificity, respectively. Adaptation of the lesion caliper detection method to account for a type of caliper specific to the case study demonstrates the intended use of BUSClean in new data distributions and improved performance in lesion caliper detection from 43.3% and 93.3% out-of-the-box to 92.1% and 92.3% sensitivity and specificity, respectively. Source code, example notebooks, and sample data are available at https://github.com/hawaii-ai/bus-cleaning.
We consider the problem of active learning for global sensitivity analysis of expensive black-box functions. Our aim is to efficiently learn the importance of different input variables, e.g., in vehicle safety experimentation, we study the impact of the thickness of various components on safety objectives. Since function evaluations are expensive, we use active learning to prioritize experimental resources where they yield the most value. We propose novel active learning acquisition functions that directly target key quantities of derivative-based global sensitivity measures (DGSMs) under Gaussian process surrogate models. We showcase the first application of active learning directly to DGSMs, and develop tractable uncertainty reduction and information gain acquisition functions for these measures. Through comprehensive evaluation on synthetic and real-world problems, our study demonstrates how these active learning acquisition strategies substantially enhance the sample efficiency of DGSM estimation, particularly with limited evaluation budgets. Our work paves the way for more efficient and accurate sensitivity analysis in various scientific and engineering applications.
Machine learning models are vulnerable to tiny adversarial input perturbations optimized to cause a very large output error. To measure this vulnerability, we need reliable methods that can find such adversarial perturbations. For image classification models, evaluation methodologies have emerged that have stood the test of time. However, we argue that in the area of semantic segmentation, a good approximation of the sensitivity to adversarial perturbations requires significantly more effort than what is currently considered satisfactory. To support this claim, we re-evaluate a number of well-known robust segmentation models in an extensive empirical study. We propose new attacks and combine them with the strongest attacks available in the literature. We also analyze the sensitivity of the models in fine detail. The results indicate that most of the state-of-the-art models have a dramatically larger sensitivity to adversarial perturbations than previously reported. We also demonstrate a size-bias: small objects are often more easily attacked, even if the large objects are robust, a phenomenon not revealed by current evaluation metrics. Our results also demonstrate that a diverse set of strong attacks is necessary, because different models are often vulnerable to different attacks.
While the biases of language models in production are extensively documented, the biases of their guardrails have been neglected. This paper studies how contextual information about the user influences the likelihood of an LLM to refuse to execute a request. By generating user biographies that offer ideological and demographic information, we find a number of biases in guardrail sensitivity on GPT-3.5. Younger, female, and Asian-American personas are more likely to trigger a refusal guardrail when requesting censored or illegal information. Guardrails are also sycophantic, refusing to comply with requests for a political position the user is likely to disagree with. We find that certain identity groups and seemingly innocuous information, e.g., sports fandom, can elicit changes in guardrail sensitivity similar to direct statements of political ideology. For each demographic category and even for American football team fandom, we find that ChatGPT appears to infer a likely political ideology and modify guardrail behavior accordingly.
Image quality assessment (IQA) has long been a fundamental challenge in image understanding. In recent years, deep learning-based IQA methods have shown promising performance. However, the lack of large amounts of labeled data in the IQA field has hindered further advancements in these methods. This paper introduces DSMix, a novel data augmentation technique specifically designed for IQA tasks, aiming to overcome this limitation. DSMix leverages the distortion-induced sensitivity map (DSM) of an image as prior knowledge. It applies cut and mix operations to diverse categories of synthetic distorted images, assigning confidence scores to class labels based on the aforementioned prior knowledge. In the pre-training phase using DSMix-augmented data, knowledge distillation is employed to enhance the model's ability to extract semantic features. Experimental results on both synthetic and authentic IQA datasets demonstrate the significant predictive and generalization performance achieved by DSMix, without requiring fine-tuning of the full model. Code is available at \url{https://github.com/I2-Multimedia-Lab/DSMix}.
Warning: This paper contains examples of stereotypes and biases. Large Language Models (LLMs) exhibit considerable social biases, and various studies have tried to evaluate and mitigate these biases accurately. Previous studies use downstream tasks as prompts to examine the degree of social biases for evaluation and mitigation. While LLMs' output highly depends on prompts, previous studies evaluating and mitigating bias have often relied on a limited variety of prompts. In this paper, we investigate the sensitivity of LLMs when changing prompt variations (task instruction and prompt, few-shot examples, debias-prompt) by analyzing task performance and social bias of LLMs. Our experimental results reveal that LLMs are highly sensitive to prompts to the extent that the ranking of LLMs fluctuates when comparing models for task performance and social bias. Additionally, we show that LLMs have tradeoffs between performance and social bias caused by the prompts. Less bias from prompt setting may result in reduced performance. Moreover, the ambiguity of instances is one of the reasons for this sensitivity to prompts in advanced LLMs, leading to various outputs. We recommend using diverse prompts, as in this study, to compare the effects of prompts on social bias in LLMs.
We present a novel method for explainable vertebral fracture assessment (XVFA) in low-dose radiographs using deep neural networks, incorporating vertebra detection and keypoint localization with uncertainty estimates. We incorporate Genant's semi-quantitative criteria as a differentiable rule-based means of classifying both vertebra fracture grade and morphology. Unlike previous work, XVFA provides explainable classifications relatable to current clinical methodology, as well as uncertainty estimations, while at the same time surpassing state-of-the art methods with a vertebra-level sensitivity of 93% and end-to-end AUC of 97% in a challenging setting. Moreover, we compare intra-reader agreement with model uncertainty estimates, with model reliability on par with human annotators.
Visual Place Recognition (VPR) plays a critical role in many localization and mapping pipelines. It consists of retrieving the closest sample to a query image, in a certain embedding space, from a database of geotagged references. The image embedding is learned to effectively describe a place despite variations in visual appearance, viewpoint, and geometric changes. In this work, we formulate how limitations in the Geographic Distance Sensitivity of current VPR embeddings result in a high probability of incorrectly sorting the top-k retrievals, negatively impacting the recall. In order to address this issue in single-stage VPR, we propose a novel mining strategy, CliqueMining, that selects positive and negative examples by sampling cliques from a graph of visually similar images. Our approach boosts the sensitivity of VPR embeddings at small distance ranges, significantly improving the state of the art on relevant benchmarks. In particular, we raise recall@1 from 75% to 82% in MSLS Challenge, and from 76% to 90% in Nordland. Models and code are available at https://github.com/serizba/cliquemining.
Differential privacy (DP) is a widely used approach for mitigating privacy risks when training machine learning models on sensitive data. DP mechanisms add noise during training to limit the risk of information leakage. The scale of the added noise is critical, as it determines the trade-off between privacy and utility. The standard practice is to select the noise scale to satisfy a given privacy budget $\varepsilon$. This privacy budget is in turn interpreted in terms of operational attack risks, such as accuracy, sensitivity, and specificity of inference attacks aimed to recover information about the training data records. We show that first calibrating the noise scale to a privacy budget $\varepsilon$, and then translating {\epsilon} to attack risk leads to overly conservative risk assessments and unnecessarily low utility. Instead, we propose methods to directly calibrate the noise scale to a desired attack risk level, bypassing the step of choosing $\varepsilon$. For a given notion of attack risk, our approach significantly decreases noise scale, leading to increased utility at the same level of privacy. We empirically demonstrate that calibrating noise to attack sensitivity/specificity, rather than $\varepsilon$, when training privacy-preserving ML models substantially improves model accuracy for the same risk level. Our work provides a principled and practical way to improve the utility of privacy-preserving ML without compromising on privacy. The code is available at https://github.com/Felipe-Gomez/riskcal
We propose a novel approach for estimating conditional or parametric expectations in the setting where obtaining samples or evaluating integrands is costly. Through the framework of probabilistic numerical methods (such as Bayesian quadrature), our novel approach allows to incorporates prior information about the integrands especially the prior smoothness knowledge about the integrands and the conditional expectation. As a result, our approach provides a way of quantifying uncertainty and leads to a fast convergence rate, which is confirmed both theoretically and empirically on challenging tasks in Bayesian sensitivity analysis, computational finance and decision making under uncertainty.
We introduce a new method to jointly reduce the dimension of the input and output space of a function between high-dimensional spaces. Choosing a reduced input subspace influences which output subspace is relevant and vice versa. Conventional methods focus on reducing either the input or output space, even though both are often reduced simultaneously in practice. Our coupled approach naturally supports goal-oriented dimension reduction, where either an input or output quantity of interest is prescribed. We consider, in particular, goal-oriented sensor placement and goal-oriented sensitivity analysis, which can be viewed as dimension reduction where the most important output or, respectively, input components are chosen. Both applications present difficult combinatorial optimization problems with expensive objectives such as the expected information gain and Sobol' indices. By optimizing gradient-based bounds, we can determine the most informative sensors and most influential parameters as the largest diagonal entries of some diagnostic matrices, thus bypassing the combinatorial optimization and objective evaluation.
Differential privacy upper-bounds the information leakage of machine learning models, yet providing meaningful privacy guarantees has proven to be challenging in practice. The private prediction setting where model outputs are privatized is being investigated as an alternate way to provide formal guarantees at prediction time. Most current private prediction algorithms, however, rely on global sensitivity for noise calibration, which often results in large amounts of noise being added to the predictions. Data-specific noise calibration, such as smooth sensitivity, could significantly reduce the amount of noise added, but were so far infeasible to compute exactly for modern machine learning models. In this work we provide a novel and practical approach based on convex relaxation and bound propagation to compute a provable upper-bound for the local and smooth sensitivity of a prediction. This bound allows us to reduce the magnitude of noise added or improve privacy accounting in the private prediction setting. We validate our framework on datasets from financial services, medical image classification, and natural language processing and across models and find our approach to reduce the noise added by up to order of magnitude.
Large Language Models (LLMs) changed the way we design and interact with software systems. Their ability to process and extract information from text has drastically improved productivity in a number of routine tasks. Developers that want to include these models in their software stack, however, face a dreadful challenge: debugging LLMs' inconsistent behavior across minor variations of the prompt. We therefore introduce two metrics for classification tasks, namely sensitivity and consistency, which are complementary to task performance. First, sensitivity measures changes of predictions across rephrasings of the prompt, and does not require access to ground truth labels. Instead, consistency measures how predictions vary across rephrasings for elements of the same class. We perform an empirical comparison of these metrics on text classification tasks, using them as guideline for understanding failure modes of the LLM. Our hope is that sensitivity and consistency will be helpful to guide prompt engineering and obtain LLMs that balance robustness with performance.