Machine Learning (cs.LG)

Abstract: Differential privacy (DP), as a rigorous mathematical definition quantifying privacy leakage, has become a well-accepted standard for privacy protection. Combined with powerful machine learning techniques, differentially private machine learning (DPML) is increasingly important. As the most classic DPML algorithm, DP-SGD incurs a significant loss of utility, which hinders DPML's deployment in practice. Many studies have recently proposed improved algorithms based on DP-SGD to mitigate utility loss. However, these studies are isolated and cannot comprehensively measure the performance of improvements proposed in algorithms. More importantly, there is a lack of comprehensive research to compare improvements in these DPML algorithms across utility, defensive capabilities, and generalizability. We fill this gap by performing a holistic measurement of improved DPML algorithms on utility and defense capability against membership inference attacks (MIAs) on image classification tasks. We first present a taxonomy of where improvements are located in the machine learning life cycle. Based on our taxonomy, we jointly perform an extensive measurement study of the improved DPML algorithms. We also cover state-of-the-art label differential privacy (Label DP) algorithms in the evaluation. According to our empirical results, DP can effectively defend against MIAs, and sensitivity-bounding techniques such as per-sample gradient clipping play an important role in defense. We also explore some improvements that can maintain model utility and defend against MIAs more effectively. Experiments show that Label DP algorithms achieve less utility loss but are fragile to MIAs. To support our evaluation, we implement a modular re-usable software, DPMLBench, which enables sensitive data owners to deploy DPML algorithms and serves as a benchmark tool for researchers and practitioners.

Abstract: Generative models have advantageous characteristics for classification tasks such as the availability of unsupervised data and calibrated confidence, whereas discriminative models have advantages in terms of the simplicity of their model structures and learning algorithms and their ability to outperform their generative counterparts. In this paper, we propose a method to train a hybrid of discriminative and generative models in a single neural network (NN), which exhibits the characteristics of both models. The key idea is the Gaussian-coupled softmax layer, which is a fully connected layer with a softmax activation function coupled with Gaussian distributions. This layer can be embedded into an NN-based classifier and allows the classifier to estimate both the class posterior distribution and the class-conditional data distribution. We demonstrate that the proposed hybrid model can be applied to semi-supervised learning and confidence calibration.

Abstract: Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demand low job completion time (JCT) for model inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long JCT. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize JCT with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate states between GPU memory and host memory for LLM inference. We build a system prototype of FastServe based on NVIDIA FasterTransformer. Experimental results show that compared to the state-of-the-art solution Orca, FastServe improves the average and tail JCT by up to 5.1$\times$ and 6.4$\times$, respectively.

Abstract: Federated Learning (FL) is a widely embraced paradigm for distilling artificial intelligence from distributed mobile data. However, the deployment of FL in mobile networks can be compromised by exposure to interference from neighboring cells or jammers. Existing interference mitigation techniques require multi-cell cooperation or at least interference channel state information, which is expensive in practice. On the other hand, power control that treats interference as noise may not be effective due to limited power budgets, and also that this mechanism can trigger countermeasures by interference sources. As a practical approach for protecting FL against interference, we propose Spectrum Breathing, which cascades stochastic-gradient pruning and spread spectrum to suppress interference without bandwidth expansion. The cost is higher learning latency by exploiting the graceful degradation of learning speed due to pruning. We synchronize the two operations such that their levels are controlled by the same parameter, Breathing Depth. To optimally control the parameter, we develop a martingale-based approach to convergence analysis of Over-the-Air FL with spectrum breathing, termed AirBreathing FL. We show a performance tradeoff between gradient-pruning and interference-induced error as regulated by the breathing depth. Given receive SIR and model size, the optimization of the tradeoff yields two schemes for controlling the breathing depth that can be either fixed or adaptive to channels and the learning process. As shown by experiments, in scenarios where traditional Over-the-Air FL fails to converge in the presence of strong interference, AirBreahing FL with either fixed or adaptive breathing depth can ensure convergence where the adaptive scheme achieves close-to-ideal performance.

Abstract: Learning causal structure among event types from discrete-time event sequences is a particularly important but challenging task. Existing methods, such as the multivariate Hawkes processes based methods, mostly boil down to learning the so-called Granger causality which assumes that the cause event happens strictly prior to its effect event. Such an assumption is often untenable beyond applications, especially when dealing with discrete-time event sequences in low-resolution; and typical discrete Hawkes processes mainly suffer from identifiability issues raised by the instantaneous effect, i.e., the causal relationship that occurred simultaneously due to the low-resolution data will not be captured by Granger causality. In this work, we propose Structure Hawkes Processes (SHPs) that leverage the instantaneous effect for learning the causal structure among events type in discrete-time event sequence. The proposed method is featured with the minorization-maximization of the likelihood function and a sparse optimization scheme. Theoretical results show that the instantaneous effect is a blessing rather than a curse, and the causal structure is identifiable under the existence of the instantaneous effect. Experiments on synthetic and real-world data verify the effectiveness of the proposed method.

Abstract: Graph Neural Networks (GNNs) are a pertinent tool for any machine learning task due to their ability to learn functions over graph structures, a powerful and expressive data representation. The detection of communities, an unsupervised task has increasingly been performed with GNNs. Clustering nodes in a graph using the multi-dimensionality of node features with the connectivity of the graph has many applications to real world tasks from social networks to genomics. Unfortunately, there is currently a gap in the literature with no established sufficient benchmarking environment for fairly and rigorously evaluating GNN based community detection, thereby potentially impeding progress in this nascent field. We observe the particular difficulties in this setting is the ambiguous hyperparameter tuning environments combined with conflicting metrics of performance and evaluation datasets. In this work, we propose and evaluate frameworks for the consistent comparisons of community detection algorithms using GNNs. With this, we show the strong dependence of the performance to the experimental settings, exacerbated by factors such as the use of GNNs and the unsupervised nature of the task, providing clear motivation for the use of a framework to facilitate congruent research in the field.

Abstract: Monotone missing data is a common problem in data analysis. However, imputation combined with dimensionality reduction can be computationally expensive, especially with the increasing size of datasets. To address this issue, we propose a Blockwise principal component analysis Imputation (BPI) framework for dimensionality reduction and imputation of monotone missing data. The framework conducts Principal Component Analysis (PCA) on the observed part of each monotone block of the data and then imputes on merging the obtained principal components using a chosen imputation technique. BPI can work with various imputation techniques and can significantly reduce imputation time compared to conducting dimensionality reduction after imputation. This makes it a practical and efficient approach for large datasets with monotone missing data. Our experiments validate the improvement in speed. In addition, our experiments also show that while applying MICE imputation directly on missing data may not yield convergence, applying BPI with MICE for the data may lead to convergence.

Abstract: Correlation matrix visualization is essential for understanding the relationships between variables in a dataset, but missing data can pose a significant challenge in estimating correlation coefficients. In this paper, we compare the effects of various missing data methods on the correlation plot, focusing on two common missing patterns: random and monotone. We aim to provide practical strategies and recommendations for researchers and practitioners in creating and analyzing the correlation plot. Our experimental results suggest that while imputation is commonly used for missing data, using imputed data for plotting the correlation matrix may lead to a significantly misleading inference of the relation between the features. We recommend using DPER, a direct parameter estimation approach, for plotting the correlation matrix based on its performance in the experiments.

Abstract: Neural network (NN) designed for challenging machine learning tasks is in general a highly nonlinear mapping that contains massive variational parameters. High complexity of NN, if unbounded or unconstrained, might unpredictably cause severe issues including over-fitting, loss of generalization power, and unbearable cost of hardware. In this work, we propose a general compression scheme that significantly reduces the variational parameters of NN by encoding them to multi-layer tensor networks (TN's) that contain exponentially-fewer free parameters. Superior compression performance of our scheme is demonstrated on several widely-recognized NN's (FC-2, LeNet-5, and VGG-16) and datasets (MNIST and CIFAR-10), surpassing the state-of-the-art method based on shallow tensor networks. For instance, about 10 million parameters in the three convolutional layers of VGG-16 are compressed in TN's with just $632$ parameters, while the testing accuracy on CIFAR-10 is surprisingly improved from $81.14\%$ by the original NN to $84.36\%$ after compression. Our work suggests TN as an exceptionally efficient mathematical structure for representing the variational parameters of NN's, which superiorly exploits the compressibility than the simple multi-way arrays.

Abstract: We study best arm identification in a variant of the multi-armed bandit problem where the learner has limited precision in arm selection. The learner can only sample arms via certain exploration bundles, which we refer to as boxes. In particular, at each sampling epoch, the learner selects a box, which in turn causes an arm to get pulled as per a box-specific probability distribution. The pulled arm and its instantaneous reward are revealed to the learner, whose goal is to find the best arm by minimising the expected stopping time, subject to an upper bound on the error probability. We present an asymptotic lower bound on the expected stopping time, which holds as the error probability vanishes. We show that the optimal allocation suggested by the lower bound is, in general, non-unique and therefore challenging to track. We propose a modified tracking-based algorithm to handle non-unique optimal allocations, and demonstrate that it is asymptotically optimal. We also present non-asymptotic lower and upper bounds on the stopping time in the simpler setting when the arms accessible from one box do not overlap with those of others.

Abstract: The success of self-supervised learning in computer vision and natural language processing has motivated pretraining methods on tabular data. However, most existing tabular self-supervised learning models fail to leverage information across multiple data tables and cannot generalize to new tables. In this work, we introduce XTab, a framework for cross-table pretraining of tabular transformers on datasets from various domains. We address the challenge of inconsistent column types and quantities among tables by utilizing independent featurizers and using federated learning to pretrain the shared component. Tested on 84 tabular prediction tasks from the OpenML-AutoML Benchmark (AMLB), we show that (1) XTab consistently boosts the generalizability, learning speed, and performance of multiple tabular transformers, (2) by pretraining FT-Transformer via XTab, we achieve superior performance than other state-of-the-art tabular deep learning models on various tasks such as regression, binary, and multiclass classification.

Abstract: Proposing an effective and flexible matrix to represent a graph is a fundamental challenge that has been explored from multiple perspectives, e.g., filtering in Graph Fourier Transforms. In this work, we develop a novel and general framework which unifies many existing GNN models from the view of parameterized decomposition and filtering, and show how it helps to enhance the flexibility of GNNs while alleviating the smoothness and amplification issues of existing models. Essentially, we show that the extensively studied spectral graph convolutions with learnable polynomial filters are constrained variants of this formulation, and releasing these constraints enables our model to express the desired decomposition and filtering simultaneously. Based on this generalized framework, we develop models that are simple in implementation but achieve significant improvements and computational efficiency on a variety of graph learning tasks. Code is available at https://github.com/qslim/PDF.

Abstract: Heart attack remain one of the greatest contributors to mortality in the United States and globally. Patients admitted to the intensive care unit (ICU) with diagnosed heart attack (myocardial infarction or MI) are at higher risk of death. In this study, we use two retrospective cohorts extracted from the eICU and MIMIC-IV databases, to develop a novel pseudo-dynamic machine learning framework for mortality prediction in the ICU with interpretability and clinical risk analysis. The method provides accurate prediction for ICU patients up to 24 hours before the event and provide time-resolved interpretability results. The performance of the framework relying on extreme gradient boosting was evaluated on a held-out test set from eICU, and externally validated on the MIMIC-IV cohort using the most important features identified by time-resolved Shapley values achieving AUCs of 91.0 (balanced accuracy of 82.3) for 6-hour prediction of mortality respectively. We show that our framework successfully leverages time-series physiological measurements by translating them into stacked static prediction problems to be robustly predictive through time in the ICU stay and can offer clinical insight from time-resolved interpretability

Abstract: Different from conventional federated learning, personalized federated learning (PFL) is able to train a customized model for each individual client according to its unique requirement. The mainstream approach is to adopt a kind of weighted aggregation method to generate personalized models, in which weights are determined by the loss value or model parameters among different clients. However, such kinds of methods require clients to download others' models. It not only sheer increases communication traffic but also potentially infringes data privacy. In this paper, we propose a new PFL algorithm called \emph{FedDWA (Federated Learning with Dynamic Weight Adjustment)} to address the above problem, which leverages the parameter server (PS) to compute personalized aggregation weights based on collected models from clients. In this way, FedDWA can capture similarities between clients with much less communication overhead. More specifically, we formulate the PFL problem as an optimization problem by minimizing the distance between personalized models and guidance models, so as to customize aggregation weights for each client. Guidance models are obtained by the local one-step ahead adaptation on individual clients. Finally, we conduct extensive experiments using five real datasets and the results demonstrate that FedDWA can significantly reduce the communication traffic and achieve much higher model accuracy than the state-of-the-art approaches.

Abstract: We show the convergence of Wasserstein inverse reinforcement learning (WIRL) for multi-objective optimizations with the projective subgradient method by formulating an inverse problem of the optimization problem that is equivalent to WIRL for multi-objective optimizations. In addition, we prove convergence of inverse reinforcement learning (maximum entropy inverse reinforcement learning, guid cost learning) for multi-objective optimization with the projective subgradient method.

Abstract: Wildfire propagation is a highly stochastic process where small changes in environmental conditions (such as wind speed and direction) can lead to large changes in observed behaviour. A traditional approach to quantify uncertainty in fire-front progression is to generate probability maps via ensembles of simulations. However, use of ensembles is typically computationally expensive, which can limit the scope of uncertainty analysis. To address this, we explore the use of a spatio-temporal neural-based modelling approach to directly estimate the likelihood of fire propagation given uncertainty in input parameters. The uncertainty is represented by deliberately perturbing the input weather forecast during model training. The computational load is concentrated in the model training process, which allows larger probability spaces to be explored during deployment. Empirical evaluations indicate that the proposed model achieves comparable fire boundaries to those produced by the traditional SPARK simulation platform, with an overall Jaccard index (similarity score) of 67.4% on a set of 35 simulated fires. When compared to a related neural model (emulator) which was employed to generate probability maps via ensembles of emulated fires, the proposed approach produces competitive Jaccard similarity scores while being approximately an order of magnitude faster.

Abstract: Graph neural networks aim to learn representations for graph-structured data and show impressive performance, particularly in node classification. Recently, many methods have studied the representations of GNNs from the perspective of optimization goals and spectral graph theory. However, the feature space that dominates representation learning has not been systematically studied in graph neural networks. In this paper, we propose to fill this gap by analyzing the feature space of both spatial and spectral models. We decompose graph neural networks into determined feature spaces and trainable weights, providing the convenience of studying the feature space explicitly using matrix space analysis. In particular, we theoretically find that the feature space tends to be linearly correlated due to repeated aggregations. Motivated by these findings, we propose 1) feature subspaces flattening and 2) structural principal components to expand the feature space. Extensive experiments verify the effectiveness of our proposed more comprehensive feature space, with comparable inference time to the baseline, and demonstrate its efficient convergence capability.

Abstract: Machine learning (ML) in healthcare presents numerous opportunities for enhancing patient care, population health, and healthcare providers' workflows. However, the real-world clinical and cost benefits remain limited due to challenges in data privacy, heterogeneous data sources, and the inability to fully leverage multiple data modalities. In this perspective paper, we introduce "patchwork learning" (PL), a novel paradigm that addresses these limitations by integrating information from disparate datasets composed of different data modalities (e.g., clinical free-text, medical images, omics) and distributed across separate and secure sites. PL allows the simultaneous utilization of complementary data sources while preserving data privacy, enabling the development of more holistic and generalizable ML models. We present the concept of patchwork learning and its current implementations in healthcare, exploring the potential opportunities and applicable data sources for addressing various healthcare challenges. PL leverages bridging modalities or overlapping feature spaces across sites to facilitate information sharing and impute missing data, thereby addressing related prediction tasks. We discuss the challenges associated with PL, many of which are shared by federated and multimodal learning, and provide recommendations for future research in this field. By offering a more comprehensive approach to healthcare data integration, patchwork learning has the potential to revolutionize the clinical applicability of ML models. This paradigm promises to strike a balance between personalization and generalizability, ultimately enhancing patient experiences, improving population health, and optimizing healthcare providers' workflows.

Abstract: Label noise widely exists in large-scale datasets and significantly degenerates the performances of deep learning algorithms. Due to the non-identifiability of the instance-dependent noise transition matrix, most existing algorithms address the problem by assuming the noisy label generation process to be independent of the instance features. Unfortunately, noisy labels in real-world applications often depend on both the true label and the features. In this work, we tackle instance-dependent label noise with a novel deep generative model that avoids explicitly modeling the noise transition matrix. Our algorithm leverages casual representation learning and simultaneously identifies the high-level content and style latent factors from the data. By exploiting the supervision information of noisy labels with structural causal models, our empirical evaluations on a wide range of synthetic and real-world instance-dependent label noise datasets demonstrate that the proposed algorithm significantly outperforms the state-of-the-art counterparts.

Abstract: Clinical diagnosis guidelines aim at specifying the steps that may lead to a diagnosis. Guidelines enable rationalizing and normalizing clinical decisions but suffer drawbacks as they are built to cover the majority of the population and may fail in guiding to the right diagnosis for patients with uncommon conditions or multiple pathologies. Moreover, their updates are long and expensive, making them unsuitable to emerging practices. Inspired by guidelines, we formulate the task of diagnosis as a sequential decision-making problem and study the use of Deep Reinforcement Learning (DRL) algorithms trained on Electronic Health Records (EHRs) to learn the optimal sequence of observations to perform in order to obtain a correct diagnosis. Because of the variety of DRL algorithms and of their sensitivity to the context, we considered several approaches and settings that we compared to each other, and to classical classifiers. We experimented on a synthetic but realistic dataset to differentially diagnose anemia and its subtypes and particularly evaluated the robustness of various approaches to noise and missing data as those are frequent in EHRs. Within the DRL algorithms, Dueling DQN with Prioritized Experience Replay, and Dueling Double DQN with Prioritized Experience Replay show the best and most stable performances. In the presence of imperfect data, the DRL algorithms show competitive, but less stable performances when compared to the classifiers (Random Forest and XGBoost); although they enable the progressive generation of a pathway to the suggested diagnosis, which can both guide or explain the decision process.

Abstract: Measuring similarity of neural networks has become an issue of great importance and research interest to understand and utilize differences of neural networks. While there are several perspectives on how neural networks can be similar, we specifically focus on two complementing perspectives, i.e., (i) representational similarity, which considers how activations of intermediate neural layers differ, and (ii) functional similarity, which considers how models differ in their outputs. In this survey, we provide a comprehensive overview of these two families of similarity measures for neural network models. In addition to providing detailed descriptions of existing measures, we summarize and discuss results on the properties and relationships of these measures, and point to open research problems. Further, we provide practical recommendations that can guide researchers as well as practitioners in applying the measures. We hope our work lays a foundation for our community to engage in more systematic research on the properties, nature and applicability of similarity measures for neural network models.

Abstract: Neural networks are a very general type of model capable of learning various relationships between multiple variables. One example of such relationships, particularly interesting in practice, is the input-output relation of nonlinear systems, which has a multitude of applications. Studying models capable of estimating such relation is a broad discipline with numerous theoretical and practical results. Neural networks are very general, but multiple special cases exist, including convolutional neural networks and recurrent neural networks, which are adjusted for specific applications, which are image and sequence processing respectively. We formulate a hypothesis that adjusting general network structure by incorporating frequency information into it should result in a network specifically well suited to nonlinear system identification. Moreover, we show that it is possible to add this frequency information without the loss of generality from a theoretical perspective. We call this new structure Frequency-Supported Neural Network (FSNN) and empirically investigate its properties.