We introduce Probe Pruning (PP), a novel framework for online, dynamic, structured pruning of Large Language Models (LLMs) applied in a batch-wise manner. PP leverages the insight that not all samples and tokens contribute equally to the model's output, and probing a small portion of each batch effectively identifies crucial weights, enabling tailored dynamic pruning for different batches. It comprises three main stages: probing, history-informed pruning, and full inference. In the probing stage, PP selects a small yet crucial set of hidden states, based on residual importance, to run a few model layers ahead. During the history-informed pruning stage, PP strategically integrates the probing states with historical states. Subsequently, it structurally prunes weights based on the integrated states and the PP importance score, a metric developed specifically to assess the importance of each weight channel in maintaining performance. In the final stage, full inference is conducted on the remaining weights. A major advantage of PP is its compatibility with existing models, as it operates without requiring additional neural network modules or fine-tuning. Comprehensive evaluations of PP on LLaMA-2/3 and OPT models reveal that even minimal probing—using just 1.5% of FLOPs—can substantially enhance the efficiency of structured pruning of LLMs. For instance, when evaluated on LLaMA-2-7B with WikiText2, PP achieves a 2.56 times lower ratio of performance degradation per unit of latency reduction compared to the state-of-the-art method at a 40\% pruning ratio.

Ensuring that generative AI systems align with human values is essential but challenging, especially when considering multiple human values and their potential trade-offs. Since human values can be personalized and dynamically change over time, the desirable levels of value alignment vary across different ethnic groups, industry sectors, and user cohorts. Within existing frameworks, it is hard to define human values and align AI systems accordingly across different directions simultaneously, such as harmlessness, helpfulness, and positiveness. To address this, we develop a novel, first-principle approach called Multi-Human-Value Alignment Palette (MAP), which navigates the alignment across multiple human values in a structured and reliable way. MAP formulates the alignment problem as an optimization task with user-defined constraints, which define human value targets. It can be efficiently solved via a primal-dual approach, which determines whether a user-defined alignment target is achievable and how to achieve it. We conduct a detailed theoretical analysis of MAP by quantifying the trade-offs between values, the sensitivity to constraints, the fundamental connection between multi-value alignment and sequential alignment, and proving that linear weighted rewards are sufficient for multi-value alignment. Extensive experiments demonstrate MAP's ability to align multiple values in a principled manner while delivering strong empirical performance across various tasks.

Retrieval-Augmented Generation (RAG) has been empirically shown to enhance the performance of large language models (LLMs) in knowledge-intensive domains such as healthcare, finance, and legal contexts. Given a query, RAG retrieves relevant documents from a corpus and integrates them into the LLMs’ generation process. In this study, we investigate the adversarial robustness of RAG, focusing specifically on examining the retrieval system. First, across 225 different setup combinations of corpus, retriever, query, and targeted information, we show that retrieval systems are vulnerable to universal poisoning attacks in medical Q&A. In such attacks, adversaries generate poisoned documents containing a broad spectrum of targeted information, such as personally identifiable information. When these poisoned documents are inserted into a corpus, they can be accurately retrieved by any users, as long as attacker-specified queries are used. To understand this vulnerability, we discovered that the deviation from the query’s embedding to that of the poisoned document tends to follow a pattern in which the high similarity between the poisoned document and the query is retained, thereby enabling precise retrieval. Based on these findings, we develop a new detection-based defense to ensure the safe use of RAG. Through extensive experiments spanning various Q&A domains, we observed that our proposed method consistently achieves excellent detection rates in nearly all cases.

Federated Learning (FL) is an approach for privacypreserving Machine Learning (ML), enabling model training across multiple clients without centralized data collection. Existing incentive solutions for traditional Federated Learning (FL) focus on individual contributions to a single global objective, neglecting the nuances of clustered personalization with multiple cluster-level models and the non-monetary incentives such as personalized model appeal for clients. In this paper, we first propose to treat incentivization and personalization as interrelated challenges and solve them with an incentive mechanism that fosters personalized learning. Additionally, current methods depend on an aggregator for client clustering, which is limited by a lack of access to clients' confidential information due to privacy constraints, leading to inaccurate clustering. To overcome this, we propose direct client involvement, allowing clients to indicate their cluster membership preferences based on data distribution and incentive-driven feedback. Our approach enhances the personalized model appeal for self-aware clients with high-quality data leading to their active and consistent participation. Our evaluation demonstrates significant improvements in test accuracy ($8-45 \%$), personalized model appeal ($3-38 \%$), and participation rates ($\mathbf{31 - 100 \%}$) over existing FL models, including those addressing data heterogeneity and personalization.

The integration of large language models (LLMs) into robotic task planning has unlocked better reasoning capabilities for complex, long-horizon workflows. However, ensuring safety in LLM-driven plans remains a critical challenge, as these models often prioritize task completion over risk mitigation. This paper introduces SAFER (Safety-Aware Framework for Execution in Robotics), a multi-LLM framework designed to embed safety awareness into robotic task planning. SAFER employs a Safety Agent that operates alongside the primary task planner, providing safety feedback. Additionally, we introduce LLM-as-a-Judge, a novel metric leveraging LLMs as evaluators to quantify safety violations within generated task plans. Our framework integrates safety feedback at multiple stages of execution, enabling real-time risk assessment, proactive error correction, and transparent safety evaluation. We also integrate a control framework using Control Barrier Functions (CBFs) to ensure safety guarantees within SAFER’s task planning. We evaluated SAFER against state-of-the-art LLM planners on complex long-horizon tasks involving heterogeneous robotic agents, demonstrating its effectiveness in reducing safety violations while maintaining task efficiency. We also verify the task planner and safety planner through actual hardware experiments involving multiple robots and a human.

Xinran Wang, Enmao Diao, Qi Le, Jie Ding, Ali Anwar. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Large language models can generate rare but catastrophic outputs, such as harmful conversations or insecure code. Existing Reinforcement Learning from Human Feedback (RLHF) typically maximizes average reward, leaving high-risk tail events insufficiently controlled. We introduce Quantile‑Guided Alignment (QA), a framework that allows users to specify desired improvements at any quantile—individually or across multiple reward dimensions—thus shifting the distribution of outputs with finer control toward safer, more desirable outcomes. The method extends standard RLHF via an augmented reward formulation that enforces quantile constraints. Experiments on conversation and code‐generation tasks show that quantile alignment significantly enhances quality at targeted tails while maintaining overall performance. The results position QA as a principled route to risk‑calibrated language models with tail‑focused alignment.

Methods in the field of quickest change detection rapidly detect in real-time a change in the data-generating distribution of an online data stream. Existing methods have been able to detect this change point when the densities of the pre-and post-change distributions are known. Recent work has extended these results to the case where the pre-and post-change distributions are known only by their score functions. This work considers the case where the pre-and post-change score functions are known only to correspond to distributions in two disjoint sets. This work selects a pair of least-favorable distributions from these sets to robustify the existing score-based quickest change detection algorithm, the properties of which are studied. This paper calculates the least-favorable distributions for specific model classes and provides methods of estimating the leastfavorable distributions for common constructions. Simulation results are provided demonstrating the performance of our robust change detection algorithm.

Federated Learning (FL) is a collaborative machine learning framework that allows multiple users to train models utilizing their local data in a distributed manner. However, considerable statistical heterogeneity in local data across devices often leads to suboptimal model performance compared with independently and identically distributed (IID) data scenarios. In this paper, we introduce DynamicFL, a new FL framework that investigates the trade-offs between global model performance and communication costs for two widely adopted FL methods: Federated Stochastic Gradient Descent (FedSGD) and Federated Averaging (FedAvg). Our approach allocates diverse communication resources to clients based on their data statistical heterogeneity, considering communication resource constraints, and attains substantial performance enhancements compared to uniform communication resource allocation. Notably, our method bridges the gap between FedSGD and FedAvg, providing a flexible framework leveraging communication heterogeneity to address statistical heterogeneity in FL. Through extensive experiments, we demonstrate that DynamicFL surpasses current state-of-the-art methods with up to a 10% increase in model accuracy, demonstrating its adaptability and effectiveness in tackling data statistical heterogeneity challenges.

Collaborations among various entities, such as companies, research labs, AI agents, and edge devices, have become increasingly crucial for achieving machine learning tasks that cannot be accomplished by a single entity alone. This is likely due to factors such as security constraints, privacy concerns, and limitations in computation resources. As a result, Collaborative Learning has been gaining momentum. However, a significant challenge in practical applications of Collaborative Learning is how to effectively incentivize multiple entities to collaborate before any collaboration occurs. In this study, we propose ICL, an architectural framework for Incentivized Collaborative Learning, and provide insights into the critical issue of when and why incentives can improve collaboration performance. We showcase the concepts of ICL to specific use cases in federated learning, assisted learning, and multi-armed bandit, corroborating with both theoretical and experimental results.

Backdoor attacks pose a significant security risk to machine learning applications due to their stealthy nature and potentially serious consequences. Such attacks involve embedding triggers within a learning model with the intention of causing malicious behavior when an active trigger is present while maintaining regular functionality without it. This paper derives a fundamental understanding of backdoor attacks that applies to both discriminative and generative models, including diffusion models and large language models. We evaluate the effectiveness of any backdoor attack incorporating a constant trigger, by establishing tight lower and upper boundaries for the performance of the compromised model on both clean and backdoor test data. The developed theory answers a series of fundamental but previously underexplored problems, including (1) what are the determining factors for a backdoor attack's success, (2) what is the direction of the most effective backdoor attack, and (3) when will a human-imperceptible trigger succeed. We demonstrate the theory by conducting experiments using benchmark datasets and state-of-the-art backdoor attack scenarios. Our code is available \href{https://github.com/KeyWgh/DemystifyBackdoor}{here}.

While the Transformer architecture has achieved remarkable success across various domains, a thorough theoretical foundation explaining its optimization dynamics is yet to be fully developed. In this study, we aim to bridge this understanding gap by answering the following two core questions: (1) Which types of Transformer architectures allow Gradient Descent (GD) to achieve guaranteed convergence? and (2) Under what initial conditions and architectural specifics does the Transformer achieve rapid convergence during training? By analyzing the loss landscape of a single Transformer layer using Softmax and Gaussian attention kernels, our work provides concrete answers to these questions. Our findings demonstrate that, with appropriate weight initialization, GD can train a Transformer model (with either kernel type) to achieve a global optimal solution, especially when the input embedding dimension is large. Nonetheless, certain scenarios highlight potential pitfalls: training a Transformer using the Softmax attention kernel may sometimes lead to suboptimal local solutions. In contrast, the Gaussian attention kernel exhibits a much favorable behavior. Our empirical study further validate the theoretical findings.

Safeguarding intellectual property and preventing potential misuse of AI-generated images are of paramount importance. This paper introduces a robust and agile plug-and-play watermark detection framework, referred to as RAW. As a departure from existing encoder-decoder methods, which incorporate fixed binary codes as watermarks within latent representations, our approach introduces learnable watermarks directly into the original image data. Subsequently, we employ a classifier that is jointly trained with the watermark to detect the presence of the watermark. The proposed framework is compatible with various generative architectures and supports on-the-fly watermark injection after training. By incorporating state-of-the-art smoothing techniques, we show that the framework also provides provable guarantees regarding the false positive rate for misclassifying a watermarked image, even in the presence of adversarial attacks targeting watermark removal. Experiments on a diverse range of images generated by state-of-the-art diffusion models demonstrate substantially improved watermark encoding speed and watermark detection performance, under adversarial attacks, while maintaining image quality. Our code is publicly available [here](https://github.com/jeremyxianx/RAWatermark).

Parabolic partial differential equations (PDEs) appear in many disciplines to model the evolution of various mathematical objects, such as probability flows, value functions in control theory, and derivative prices in finance. It is often necessary to compute the solutions or a function of the solutions to a parametric PDE in multiple scenarios corresponding to different parameters of this PDE. This process often requires resolving the PDEs from scratch, which is time-consuming. To better employ existing simulations for the PDEs, we propose a framework for finding solutions to parabolic PDEs across different scenarios by meta-learning an underlying base distribution.%tasks. We build upon this base distribution to propose a method for computing solutions to parametric PDEs under different parameter settings. Finally, we illustrate the application of the proposed methods through extensive experiments in generative modeling, stochastic control, and finance. The empirical results suggest that the proposed approach improves generalization to solving new PDEs.

Classical quickest change detection algorithms require modeling pre-change and post-change distributions. Such an approach may not be feasible for various machine learning models because of the complexity of computing the explicit distributions. Additionally, these methods may suffer from a lack of robustness to model mismatch and noise. This paper develops a new variant of the classical Cumulative Sum (CUSUM) algorithm for the quickest change detection. This variant is based on Fisher divergence and the Hyvärinen score and is called the Hyvärinen score-based CUSUM (SCUSUM) algorithm. The SCUSUM algorithm allows the applications of change detection for unnormalized statistical models, i.e., models for which the probability density function contains an unknown normalization constant. The asymptotic optimality of the proposed algorithm is investigated by deriving expressions for average detection delay and the mean running time to a false alarm. Numerical results are provided to demonstrate the performance of the proposed algorithm.

Classical change detection algorithms typically require modeling pre-change and post-change distributions. The calculations may not be feasible for various machine learning models because of the complexity of computing the partition functions and normalized distributions. Additionally, these methods may suffer from a lack of robustness to model mismatch and noise. In this paper, we develop a new variant of the classical Cumulative Sum (CUSUM) change detection, namely Score-based CUSUM (SCUSUM), based on Fisher divergence and the Hyvärinen score. Our method allows the applications of the quickest change detection for unnormalized distributions. We provide a theoretical analysis of the detection delay given the constraints on false alarms. We prove the asymptotic optimality of the proposed method in some particular cases. We also provide numerical experiments to demonstrate our method’s computation, performance, and robustness advantages.

The deployment of Federated Learning (FL) systems poses various challenges such as data heterogeneity and communication efficiency. We focus on a practical FL setup that has recently drawn attention, where the data distribution on each device is not static but dynamically evolves over time. This setup, referred to as Continual Federated Learning (CFL), suffers from catastrophic forgetting, i.e., the undesired forgetting of previous knowledge after learning on new data, an issue not encountered with vanilla FL. In this work, we formally quantify catastrophic forgetting in a CFL setup, establish links to training optimization and evaluate different episodic replay approaches for CFL on a large scale real-world NLP dataset. To the best of our knowledge, this is the first such study of episodic replay for CFL. We show that storing a small set of past data boosts performance and significantly reduce forgetting, providing evidence that carefully designed sampling strategies can lead to further improvements.

A poisoning backdoor attack is a rising security concern for deep learning. This type of attack can result in the backdoored model functioning normally most of the time but exhibiting abnormal behavior when presented with inputs containing the backdoor trigger, making it difficult to detect and prevent. In this work, we propose the adaptability hypothesis to understand when and why a backdoor attack works for general learning models, including deep neural networks, based on the theoretical investigation of classical kernel-based learning models. The adaptability hypothesis postulates that for an effective attack, the effect of incorporating a new dataset on the predictions of the original data points will be small, provided that the original data points are distant from the new dataset. Experiments on benchmark image datasets and state-of-the-art backdoor attacks for deep neural networks are conducted to corroborate the hypothesis. Our finding provides insight into the factors that affect the attack’s effectiveness and has implications for the design of future attacks and defenses.

Unsupervised domain adaptation (UDA) is a popular machine learning technique that allows one to train models over diverse data collected from different domains. However, this technique requires the learner to collect a large number of properly labeled data samples, which can be costly and unrealistic in many applications. In this work, we propose a decentralized assisted learning framework for UDA. In this framework, a learner has only a limited number of labeled data samples collected from a certain source domain and aims to train a classifier for the target domain. To improve domain adaptation performance, it seeks assistance by interacting with an external service provider, who possesses many labeled data samples collected from a related source domain. We develop an assisted UDA algorithm that avoids data sharing and can significantly improve the learner’s domain adaptation performance within a few rounds of interaction. Experiments using deep neural networks on benchmark datasets demonstrate the effectiveness of this algorithm.

Backdoor attacks involve inserting poisoned samples during training, resulting in a model containing a hidden backdoor that can trigger specific behaviors without impacting performance on normal samples. These attacks are challenging to detect, as the backdoored model appears normal until activated by the backdoor trigger, rendering them particularly stealthy. In this study, we devise a unified inference-stage detection framework to defend against backdoor attacks. We first rigorously formulate the inference-stage backdoor detection problem, encompassing various existing methods, and discuss several challenges and limitations. We then propose a framework with provable guarantees on the false positive rate or the probability of misclassifying a clean sample. Further, we derive the most powerful detection rule to maximize the detection power, namely the rate of accurately identifying a backdoor sample, given a false positive rate under classical learning scenarios. Based on the theoretically optimal detection rule, we suggest a practical and effective approach for real-world applications based on the latent representations of backdoored deep nets. We extensively evaluate our method on 14 different backdoor attacks using Computer Vision (CV) and Natural Language Processing (NLP) benchmark datasets. The experimental findings align with our theoretical results. We significantly surpass the state-of-the-art methods, e.g., up to 300\% improvement on the detection power as evaluated by AUCROC, over the state-of-the-art defense against advanced adaptive backdoor attacks.

Detecting an abrupt and persistent change in the underlying distribution of online data streams is an important problem in many applications. This paper proposes a new robust score-based algorithm called RSCUSUM, which can be applied to unnormalized models and addresses the issue of unknown post-change distributions. RSCUSUM replaces the Kullback-Leibler divergence with the Fisher divergence between pre- and post-change distributions for computational efficiency in unnormalized statistical models and introduces a notion of the “least favorable” distribution for robust change detection. The algorithm and its theoretical analysis are demonstrated through simulation studies.

LASSO regularization is a popular regression tool to enhance the prediction accuracy of statistical models by performing variable selection through the$\ell _{1}$penalty, initially formulated for the linear model and its variants. In this paper, the territory of LASSO is extended to two-layer ReLU neural networks, a fashionable and powerful nonlinear regression model. Specifically, given a neural network whose output$y$depends only on a small subset of input$\boldsymbol {x}$, denoted by$\mathcal {S}^{\star }$, we prove that the LASSO estimator can stably reconstruct the neural network and identify$\mathcal {S}^{\star }$when the number of samples scales logarithmically with the input dimension. This challenging regime has been well understood for linear models while barely studied for neural networks. Our theory lies in an extended Restricted Isometry Property (RIP)-based analysis framework for two-layer ReLU neural networks, which may be of independent interest to other LASSO or neural network settings. Based on the result, we advocate a neural network-based variable selection method. Experiments on simulated and real-world datasets show promising performance of the variable selection approach compared with existing techniques.

We propose a new structured pruning framework for compressing Deep Neural Networks (DNNs) with skip-connections, based on measuring the statistical dependency of hidden layers and predicted outputs. The dependence measure defined by the energy statistics of hidden layers serves as a model-free measure of information between the feature maps and the output of the network. The estimated dependence measure is subsequently used to prune a collection of redundant and uninformative layers. Extensive numerical experiments on various architectures show the efficacy of the proposed pruning approach with competitive performance to state-of-the-art methods.

Federated learning (FL) has been developed as a promising framework to leverage the resources of edge devices, enhance customers’ privacy, comply with regulations, and reduce development costs. Although many methods and applications have been developed for FL, several critical challenges for practical FL systems remain unaddressed. This paper provides an outlook on FL development as part of the ICASSP 2022 special session entitled "Frontiers of Federated Learning: Applications, Challenges, and Opportunities." The outlook is categorized into five emerging directions of FL, namely algorithm foundation, personalization, hardware and security constraints, lifelong learning, and nonstandard data. Our unique perspectives are backed by practical observations from large-scale federated systems for edge devices.

Supervised learning scenarios, where labels and features are possibly mismatched, have been an emerging concern in machine learning applications. For example, researchers often need to align heterogeneous data from multiple resources to the same entities without a unique identifier in the socioeconomic study. Such a mismatch problem can significantly affect the learning performance if it is not appropriately addressed. Due to the combinatorial nature of the mismatch problem, existing methods are often designed for small datasets and simple linear models but are not scalable to large-scale datasets and complex models. In this paper, we first present a new formulation of the mismatch problem that supports continuous optimization problems and allows for gradient-based methods. Moreover, we develop a computation and memory efficient method to process complex data and models. Empirical studies on synthetic and real-world data show significantly better performance of the proposed algorithms than state-of-the-art methods.

Federated Learning (FL) is an efficient distributed machine learning algorithm that promises to reduce data migration costs to a centralized repository, alleviate regulatory data restrictions and maintain data privacy. However, it suffers from data heterogeneity, i.e., data distribution across silos is often non-identical and independent (non-iid), making optimization difficult. Further, a significant but rarely studied challenge in FL is the lack of annotated data for training. This challenge is more pronounced in the medical field since data annotations require precision and are highly labor-intensive and time-consuming. That is why a few sites have minimal or no annotated data, which wastes valuable data resources for ML training. In this work, we investigate these two challenges of Federated Learning on a publicly available realistic federated medical dataset, KiTS19. First, we explore Federated Learning for Tumor Segmentation task on the Federated version of the KiTS19 dataset for the first time. We show that FL can maintain 96% of model accuracy compared to the centralized model accuracy with ten institution collaboration. In addition, we investigate the benefits of transfer learning to address the challenge of data heterogeneity and show that 5% accuracy improvement is achieved by using a pre-trained model in FL. Moreover, we propose a Federated semi-supervised learning (FSSL) framework to address the challenge of the lack of annotations at some silos. We show that unlabelled silos add 11% to the model’s efficiency compared with the model trained on labeled silos alone.

In the context of personalized federated learning (FL), the critical challenge is to balance local model improvement and global model tuning when the personal and global objectives may not be exactly aligned. Inspired by Bayesian hierarchical models, we develop a self-aware personalized FL method where each client can automatically balance the training of its local personal model and the global model that implicitly contributes to other clients' training. Such a balance is derived from the inter-client and intra-client uncertainty quantification. A larger inter-client variation implies more personalization is needed. Correspondingly, our method uses uncertainty-driven local training steps an aggregation rule instead of conventional local fine-tuning and sample size-based aggregation. With experimental studies on synthetic data, Amazon Alexa audio data, and public datasets such as MNIST, FEMNIST, CIFAR10, and Sent140, we show that our proposed method can achieve significantly improved personalization performance compared with the existing counterparts.

Collaborations among multiple organizations, such as financial institutions, medical centers, and retail markets in decentralized settings are crucial to providing improved service and performance. However, the underlying organizations may have little interest in sharing their local data, models, and objective functions. These requirements have created new challenges for multi-organization collaboration. In this work, we propose Gradient Assisted Learning (GAL), a new method for multiple organizations to assist each other in supervised learning tasks without sharing local data, models, and objective functions. In this framework, all participants collaboratively optimize the aggregate of local loss functions, and each participant autonomously builds its own model by iteratively fitting the gradients of the overarching objective function. We also provide asymptotic convergence analysis and practical case studies of GAL. Experimental studies demonstrate that GAL can achieve performance close to centralized learning when all data, models, and objective functions are fully disclosed.

Federated Learning allows the training of machine learning models by using the computation and private data resources of many distributed clients. Most existing results on Federated Learning (FL) assume the clients have ground-truth labels. However, in many practical scenarios, clients may be unable to label task-specific data due to a lack of expertise or resource. We propose SemiFL to address the problem of combining communication-efficient FL such as FedAvg with Semi-Supervised Learning (SSL). In SemiFL, clients have completely unlabeled data and can train multiple local epochs to reduce communication costs, while the server has a small amount of labeled data. We provide a theoretical understanding of the success of data augmentation-based SSL methods to illustrate the bottleneck of a vanilla combination of communication-efficient FL with SSL. To address this issue, we propose alternate training to 'fine-tune global model with labeled data' and 'generate pseudo-labels with the global model.' We conduct extensive experiments and demonstrate that our approach significantly improves the performance of a labeled server with unlabeled clients training with multiple local epochs. Moreover, our method outperforms many existing SSFL baselines and performs competitively with the state-of-the-art FL and SSL results.

A crucial problem of neural networks is to select an architecture that strikes appropriate tradeoffs between underfitting and overfitting. This work shows that$\ell _1$regularizations for two-layer neural networks can control the generalization error and sparsify the input dimension. In particular, with an appropriate$\ell _1$regularization on the output layer, the network can produce a tight statistical risk. Moreover, an appropriate$\ell _1$regularization on the input layer leads to a risk bound that does not involve the input data dimension. The results also indicate that training a wide neural network with a suitable regularization provides an alternative bias-variance tradeoff to selecting from a candidate set of neural networks. Our analysis is based on a new integration of dimension-based and norm-based complexity analysis to bound the generalization error.

It has been conjectured that the Fisher divergence is more robust to model uncertainty than the conventional Kullback-Leibler (KL) divergence. This motivates the design of a new class of robust generative auto-encoders (AE) referred to as Fisher auto-encoders. Our approach is to design Fisher AEs by minimizing the Fisher divergence between the intractable joint distribution of observed data and latent variables, with that of the postulated/modeled joint distribution. In contrast to KL-based variational AEs (VAEs), the Fisher AE can exactly quantify the distance between the true and the model-based posterior distributions. Qualitative and quantitative results are provided on both MNIST and celebA datasets demonstrating the competitive performance of Fisher AEs in terms of robustness compared to other AEs such as VAEs and Wasserstein AEs.

In this paper, we propose a new structural technique for pruning deep neural networks with skip-connections. Our approach is based on measuring the importance of feature maps in predicting the output of the model using their Fisher scores. These scores subsequently used for removing the less informative layers from the graph of the network. Extensive experiments on the classification of CIFAR-10, CIFAR-100, and SVHN data sets demonstrate the efficacy of our compressing method both in the number of parameters and operations.

The rapid development in data collecting devices and computation platforms produces an emerging number of agents, each equipped with a unique data modality over a particular population of subjects. While an agent’s predictive performance may be enhanced by transmitting others’ data to it, this is often unrealistic due to intractable transmission costs and security concerns. In this paper, we propose a method named ASCII for an agent to improve its classification performance through assistance from other agents, without sharing proprietary data and model information. The main idea is to iteratively interchange an ignorance value between 0 and 1 for each collated sample among agents, where the value represents the urgency of further assistance needed. The method is naturally suitable for privacy-aware, transmission-economical, and decentralized learning scenarios. The method is also general as it allows the agents to use arbitrary classifiers such as logistic regression, ensemble tree, and neural network, and they may be heterogeneous among agents. We demonstrate the proposed method with extensive experimental studies.

We consider the distributed learning setting where each agent or learner holds a specific parametric model and a data source. The goal is to integrate information across a set of learners and data sources to enhance the prediction accuracy of a given learner. A natural way to integrate information is to build a joint model across a group of learners that shares common parameters of interest. However, the underlying parameter sharing patterns across a set of learners may not be known a priori. Misspecifying the parameter sharing patterns or the parametric model for each learner often yields a biased estimator that degrades the prediction accuracy. We propose a general method to integrate information across a set of learners that is robust against misspecification of both models and parameter sharing patterns. The main crux of our proposed method is to sequentially incorporate additional learners that can enhance the prediction accuracy of an existing joint model based on user- specified parameter sharing patterns across a set of learners. Theoretically, we show that the proposed method can data-adaptively select a parameter sharing pattern that enhances the predictive performance of a given learner. Extensive numerical studies are conducted to assess the performance of the proposed method.

A central issue of many statistical learning problems is to select an appropriate model from a set of candidate models. Large models tend to inflate the variance (or overfitting), while small models tend to cause biases (or underfitting) for a given fixed dataset. In this work, we address the critical challenge of model selection to strike a balance between model fitting and model complexity, thus gaining reliable predictive power. We consider the task of approaching the theoretical limit of statistical learning, meaning that the selected model has the predictive performance that is as good as the best possible model given a class of potentially misspecified candidate models. We propose a generalized notion of Takeuchi's information criterion and prove that the proposed method can asymptotically achieve the optimal out-sample prediction loss under reasonable assumptions. It is the first proof of the asymptotic property of Takeuchi's information criterion to our best knowledge. Our proof applies to a wide variety of nonlinear models, loss functions, and high dimensionality (in the sense that the models' complexity can grow with sample size). The proposed method can be used as a computationally efficient surrogate for leave-one-out cross-validation. Moreover, for modeling streaming data, we propose an online algorithm that sequentially expands the model complexity to enhance selection stability and reduce computation cost. Experimental studies show that the proposed method has desirable predictive power and significantly less computational cost than some popular methods.

We present a new concept of privacy and corresponding mechanisms for privatizing data that will be collected for further learning. The privacy, named as Interval Privacy, enforces the distribution of the raw data conditional on privatized data to be the same as its unconditional distribution over a nontrivial support set. The proposed privatizing mechanism is based on interval censoring techniques, where a set of points is recorded as a set of random intervals containing them. We study some theoretical properties of the proposed privacy mechanism. We demonstrate its use with various examples. Particularly, in the context of supervised regression, we develop a general method that can adapt existing regression algorithms to address interval-valued data.

Several modern applications involve forecasting time series data that exhibit both short-time dynamics and long-time seasonality. In particular, modeling time series with multiple seasonality is a challenging task with relatively few discussions. In this paper, we propose a two-stage method for predicting time series with multi-seasonality, which does not require predetermined seasonality periods. In the first stage, we generalize the classical seasonal autoregressive moving average model to multi-seasonality scenarios. In the second stage, we utilize an appropriate criterion for lag order selection. Simulation and empirical studies show the excellent predictive performance of our method, especially when compared with a recently popular `Facebook Prophet' model for time series.

We propose a new architecture for distributed image compression from a group of distributed data sources. The work is motivated by practical needs of data-driven codec design, low power consumption, robustness, and data privacy. The proposed architecture, which we refer to as Distributed Recurrent Autoencoder for Scalable Image Compression (DRASIC), is able to train distributed encoders and one joint decoder on correlated data sources. Its compression capability is much better than the method of training codecs separately. Meanwhile, the performance of our distributed system with 10 distributed sources is only within 2 dB peak signal-to-noise ratio (PSNR) of the performance of a single codec trained with all data sources. We experiment distributed sources with different correlations and show how our data-driven methodology well matches the Slepian-Wolf Theorem in Distributed Source Coding (DSC). To the best of our knowledge, this is the first data-driven DSC framework for general distributed code design with deep learning.

Motivated by the ever-increasing demands for limited communication bandwidth and low-power consumption, we propose a new methodology, named joint Variational Autoencoders with Bernoulli mixture models (VAB), for performing clustering in the compressed data domain. The idea is to reduce the data dimension by Variational Autoencoders (VAEs) and group data representations by Bernoulli mixture models (BMMs). Once jointly trained for compression and clustering, the model can be decomposed into two parts: a data vendor that encodes the raw data into compressed data, and a data consumer that classifies the received (compressed) data. In this way, the data vendor benefits from data security and communication bandwidth, while the data consumer benefits from low computational complexity. To enable training using the gradient descent algorithm, we propose to use the Gumbel-Softmax distribution to resolve the infeasibility of the back-propagation algorithm when assessing categorical samples.

In this work, we introduce a new procedure for applying Restricted Boltzmann Machines (RBMs) to missing data inference tasks, based on linearization of the effective energy function governing the distribution of observations. We compare the performance of our proposed procedure with those obtained using existing reconstruction procedures trained on incomplete data. We place these performance comparisons within the context of the perception-distortion trade-off observed in other data reconstruction tasks, which has, until now, remained unexplored in tasks relying on incomplete training data.

Classical supervised classification tasks search for a nonlinear mapping that maps each encoded feature directly to a probability mass over the labels. Such a learning framework typically lacks the intuition that encoded features from the same class tend to be similar and thus has little interpretability for the learned features. In this paper, we propose a novel supervised learning model named Supervised-Encoding Quantizer (SEQ). The SEQ applies a quantizer to cluster and classify the encoded features. We found that the quantizer provides an interpretable graph where each cluster in the graph represents a class of data samples that have a particular style. We also trained a decoder that can decode convex combinations of the encoded features from similar and different clusters and provide guidance on style transfer between sub-classes.

Speech Emotion Recognition (SER) has emerged as a critical component of the next generation of human-machine interfacing technologies. In this work, we propose a new dual-level model that predicts emotions based on both MFCC features and mel-spectrograms produced from raw audio signals. Each utterance is preprocessed into MFCC features and two mel-spectrograms at different time-frequency resolutions. A standard LSTM processes the MFCC features, while a novel LSTM architecture, denoted as Dual-Sequence LSTM (DS-LSTM), processes the two mel-spectrograms simultaneously. The outputs are later averaged to produce a final classification of the utterance. Our proposed model achieves, on average, a weighted accuracy of 72.7% and an unweighted accuracy of 73.3% - a 6% improvement over current state-of-the-art unimodal models - and is comparable with multimodal models that leverage textual information as well as audio signals.

A technique for compressing deep neural models achieving competitive performance to state-of-the-art methods is proposed. The approach utilizes the mutual information between the feature maps and the output of the model in order to prune the redundant layers of the network. Extensive numerical experiments on both CIFAR-10, CIFAR-100, and Tiny ImageNet data sets demonstrate that the proposed method can be effective in compressing deep models, both in terms of the numbers of parameters and operations. For instance, by applying the proposed approach to DenseNet model with 0.77 million parameters and 293 million operations for classification of CIFAR-10 data set, a reduction of 62.66% and 41.00% in the number of parameters and the number of operations are respectively achieved, while increasing the test error only by less than 1%.

In an increasing number of AI scenarios, collaborations among different organizations or agents (e.g., human and robots, mobile units) are often essential to accomplish an organization-specific mission. However, to avoid leaking useful and possibly proprietary information, organizations typically enforce stringent security constraints on sharing modeling algorithms and data, which significantly limits collaborations. In this work, we introduce the Assisted Learning framework for organizations to assist each other in supervised learning tasks without revealing any organization's algorithm, data, or even task. An organization seeks assistance by broadcasting task-specific but nonsensitive statistics and incorporating others' feedback in one or more iterations to eventually improve its predictive performance. Theoretical and experimental studies, including real-world medical benchmarks, show that Assisted Learning can often achieve near-oracle learning performance as if data and training processes were centralized.

Recurrent Neural Network (RNN) and its variations such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), have become standard building blocks for learning online data of sequential nature in many research areas, including natural language processing and speech data analysis. In this paper, we present a new methodology to significantly reduce the number of parameters in RNNs while maintaining performance that is comparable or even better than classical RNNs. The new proposal, referred to as Restricted Recurrent Neural Network (RRNN), restricts the weight matrices corresponding to the input data and hidden states at each time step to share a large proportion of parameters. The new architecture can be regarded as a compression of its classical counterpart, but it does not require pre-training or sophisticated parameter fine-tuning, both of which are major issues in most existing compression techniques. Experiments on natural language modeling show that compared with its classical counterpart, the restricted recurrent architecture generally produces comparable results at about 50% compression rate. In particular, the Restricted LSTM can outperform classical RNN with even less number of parameters.

Motivated by Fisher divergence, in this paper we present a new set of information quantities which we refer to as gradient information. These measures serve as surrogates for classical information measures such as those based on logarithmic loss, Kullback-Leibler divergence, directed Shannon information, etc. in many data-processing scenarios of interest, and often provide significant computational advantage, improved stability and robustness. As an example, we apply these measures to the Chow-Liu tree algorithm, and demonstrate remarkable performance and significant computational reduction using both synthetic and real data.

We develop a methodology (referred to as kinetic prediction) for predicting time series undergoing unknown changes in their data generating distributions. Based on Kolmogorov-Tikhomirov's ε-entropy, we propose a concept called ε-predictability that quantifies the size of a model class (which can be parametric or nonparametric) and the maximal number of abrupt structural changes that guarantee the achievability of asymptotically optimal prediction. Moreover, for parametric distribution families, we extend the aforementioned kinetic prediction with discretized function spaces to its counterpart with continuous function spaces, and propose a sequential Monte Carlo-based implementation. We also extend our methodology for predicting smoothly varying data generating distributions. Under reasonable assumptions, we prove that the average predictive performance converges almost surely to the oracle bound, which corresponds to the case that the data generating distributions are known in advance. The results also shed some light on the so called “prediction-inference dilemma.” Various examples and numerical results are provided to demonstrate the wide applicability of our methodology.

Machine learning systems learn from and make predictions by building models from observed data. Because large models tend to overfit while small models tend to underfit for a given fixed dataset, a critical challenge is to select an appropriate model (e.g. set of variables/features). Model selection aims to strike a balance between the goodness of fit and model complexity, and thus to gain reliable predictive power. In this paper, we study a penalized model selection technique that asymptotically achieves the optimal expected prediction loss (referred to as the limit of learning) offered by a set of candidate models. We prove that the proposed procedure is both statistically efficient in the sense that it asymptotically approaches the limit of learning, and computationally efficient in the sense that it can be much faster than cross validation methods. Our theory applies for a wide variety of model classes, loss functions, and high dimensions (in the sense that the models' complexity can grow with data size). We released a python package with our proposed method for general usage like logistic regression and neural networks.

In this work, we propose a new inference procedure for understanding non-stationary processes, under the framework of evolutionary spectra developed by Priestley. Among various frameworks of modeling non-stationary processes, the distinguishing feature of the evolutionary spectra is its focus on the physical meaning of frequency. The classical estimate of the evolutionary spectral density is based on a double-window technique consisting of a short-Fourier transform and a smoothing. However, smoothing is known to suffer from the so-called bias leakage problem. By incorporating Thomson's multitaper method that was originally designed for stationary processes, we propose an improved estimate of the evolutionary spectral density, and analyze its bias/variance/resolution tradeoff. As an application of the new estimate, we further propose a non-parametric rank-based stationarity test, and provide various experimental studies.

To address order selection for an autoregressive model fitted to time series data, we propose a new information criterion. It has the benefits of the two well-known model selection techniques: the Akaike information criterion and the Bayesian information criterion. When the data are generated from a finite-order autoregression, the Bayesian information criterion is known to be consistent, and so is the new criterion. When the true order is infinity or suitably high with respect to the sample size, the Akaike information criterion is known to be efficient in the sense that its predictive performance is asymptotically equivalent to the best offered by the candidate models; in this case, the new criterion behaves in a similar manner. Different from the two classical criteria, the proposed criterion adaptively achieves either consistency or efficiency depending on the underlying true model. In practice, where the observed time series is given without any prior information about the model specification, the proposed order selection criterion is more flexible and reliable compared with classical approaches. Numerical results are presented, demonstrating the adaptivity of the proposed technique when applied to various data sets.

We derive the tightest known bounds on η = 2ν, where ν is the growth rate of the logarithm of the number of independent sets on a hexagonal lattice. To obtain these bounds, we generalize a method proposed by Calkin and Wilf. Their original strategy cannot immediately be used to derive bounds for η, due to the difference in symmetry between square and hexagonal lattices, so we propose a modified method and an algorithm to derive rigorous bounds on η. In particular, we prove that 1.546440708536001 ≤ η ≤ 1.5513, which improves upon the best known bounds of 1.5463 ≤ η ≤ 1.5527 given by Nagy and Zeger. Our lower bound matches the numerical estimate of Baxter up to 9 digits after the decimal point, and our upper bound can be further improved by following our method.