Personalized federated learning (PFL) is a popular framework that allows clients to have different models to address application scenarios where clients' data are in different domains. The typical model of a client in PFL features a global encoder trained by all clients to extract universal features from the raw data and personalized layers (e.g., a classifier) trained using the client's local data. Nonetheless, due to the differences between the data distributions of different clients (aka, domain gaps), the universal features produced by the global encoder largely encompass numerous components irrelevant to a certain client's local task. Some recent PFL methods address the above problem by personalizing specific parameters within the encoder. However, these methods encounter substantial challenges attributed to the high dimensionality and non-linearity of neural network parameter space. In contrast, the feature space exhibits a lower dimensionality, providing greater intuitiveness and interpretability as compared to the parameter space. To this end, we propose a novel PFL framework named FedPick. FedPick achieves PFL in the low-dimensional feature space by selecting task-relevant features adaptively for each client from the features generated by the global encoder based on its local data distribution. It presents a more accessible and interpretable implementation of PFL compared to those methods working in the parameter space. Extensive experimental results show that FedPick could effectively select task-relevant features for each client and improve model performance in cross-domain FL.
Federated learning (FL) is a new paradigm for distributed machine learning that allows a global model to be trained across multiple clients without compromising their privacy. Although FL has demonstrated remarkable success in various scenarios, recent studies mainly utilize shallow and small neural networks. In our research, we discover a significant performance decline when applying the existing FL framework to deeper neural networks, even when client data are independently and identically distributed (i.i.d.). Our further investigation shows that the decline is due to the continuous accumulation of dissimilarities among client models during the layer-by-layer back-propagation process, which we refer to as "divergence accumulation." As deeper models involve a longer chain of divergence accumulation, they tend to manifest greater divergence, subsequently leading to performance decline. Both theoretical derivations and empirical evidence are proposed to support the existence of divergence accumulation and its amplified effects in deeper models. To address this issue, we propose several technical guidelines based on reducing divergence, such as using wider models and reducing the receptive field. These approaches can greatly improve the accuracy of FL on deeper models. For example, the application of these guidelines can boost the ResNet101 model's performance by as much as 43\% on the Tiny-ImageNet dataset.
Machine learning algorithms play an important role in a variety of important decision-making processes, including targeted advertisement displays, home loan approvals, and criminal behavior predictions. Given the far-reaching impact of these algorithms, it is crucial that they operate fairly, free from bias or prejudice towards certain groups in the population. Ensuring impartiality in these algorithms is essential for promoting equality and avoiding discrimination. To this end we introduce a unified framework for randomized subset selection that incorporates group fairness constraints. Our problem involves a global utility function and a set of group utility functions for each group, here a group refers to a group of individuals (e.g., people) sharing the same attributes (e.g., gender). Our aim is to generate a distribution across feasible subsets, specifying the selection probability of each feasible set, to maximize the global utility function while meeting a predetermined quota for each group utility function in expectation. Note that there may not necessarily be any direct connections between the global utility function and each group utility function. We demonstrate that this framework unifies and generalizes many significant applications in machine learning and operations research. Our algorithmic results either improves the best known result or provide the first approximation algorithms for new applications.
Submodular function optimization has numerous applications in machine learning and data analysis, including data summarization which aims to identify a concise and diverse set of data points from a large dataset. It is important to implement fairness-aware algorithms when dealing with data items that may contain sensitive attributes like race or gender, to prevent biases that could lead to unequal representation of different groups. With this in mind, we investigate the problem of maximizing a monotone submodular function while meeting group fairness constraints. Unlike previous studies in this area, we allow for randomized solutions, with the objective being to calculate a distribution over feasible sets such that the expected number of items selected from each group is subject to constraints in the form of upper and lower thresholds, ensuring that the representation of each group remains balanced in the long term. Here a set is considered feasible if its size does not exceed a constant value of $b$. Our research includes the development of a series of approximation algorithms for this problem.
Many large vision models have been deployed on the cloud for real-time services. Meanwhile, fresh samples are continuously generated on the served mobile device. How to leverage the device-side samples to improve the cloud-side large model becomes a practical requirement, but falls into the dilemma of no raw sample up-link and no large model down-link. Specifically, the user may opt out of sharing raw samples with the cloud due to the concern of privacy or communication overhead, while the size of some large vision models far exceeds the mobile device's runtime capacity. In this work, we propose a device-cloud collaborative controlled learning framework, called DC-CCL, enabling a cloud-side large vision model that cannot be directly deployed on the mobile device to still benefit from the device-side local samples. In particular, DC-CCL vertically splits the base model into two submodels, one large submodel for learning from the cloud-side samples and the other small submodel for learning from the device-side samples and performing device-cloud knowledge fusion. Nevertheless, on-device training of the small submodel requires the output of the cloud-side large submodel to compute the desired gradients. DC-CCL thus introduces a light-weight model to mimic the large cloud-side submodel with knowledge distillation, which can be offloaded to the mobile device to control its small submodel's optimization direction. Given the decoupling nature of two submodels in collaborative learning, DC-CCL also allows the cloud to take a pre-trained model and the mobile device to take another model with a different backbone architecture.
In recent years, large amounts of electronic health records (EHRs) concerning chronic diseases, such as cancer, diabetes, and mental disease, have been collected to facilitate medical diagnosis. Modeling the dynamic properties of EHRs related to chronic diseases can be efficiently done using dynamic treatment regimes (DTRs), which are a set of sequential decision rules. While Reinforcement learning (RL) is a widely used method for creating DTRs, there is ongoing research in developing RL algorithms that can effectively handle large amounts of data. In this paper, we present a novel approach, a distributed Q-learning algorithm, for generating DTRs. The novelties of our research are as follows: 1) From a methodological perspective, we present a novel and scalable approach for generating DTRs by combining distributed learning with Q-learning. The proposed approach is specifically designed to handle large amounts of data and effectively generate DTRs. 2) From a theoretical standpoint, we provide generalization error bounds for the proposed distributed Q-learning algorithm, which are derived within the framework of statistical learning theory. These bounds quantify the relationships between sample size, prediction accuracy, and computational burden, providing insights into the performance of the algorithm. 3) From an applied perspective, we demonstrate the effectiveness of our proposed distributed Q-learning algorithm for DTRs by applying it to clinical cancer treatments. The results show that our algorithm outperforms both traditional linear Q-learning and commonly used deep Q-learning in terms of both prediction accuracy and computation cost.
Maximizing a submodular function has a wide range of applications in machine learning and data mining. One such application is data summarization whose goal is to select a small set of representative and diverse data items from a large dataset. However, data items might have sensitive attributes such as race or gender, in this setting, it is important to design \emph{fairness-aware} algorithms to mitigate potential algorithmic bias that may cause over- or under- representation of particular groups. Motivated by that, we propose and study the classic non-monotone submodular maximization problem subject to novel group fairness constraints. Our goal is to select a set of items that maximizes a non-monotone submodular function, while ensuring that the number of selected items from each group is proportionate to its size, to the extent specified by the decision maker. We develop the first constant-factor approximation algorithms for this problem. We also extend the basic model to incorporate an additional global size constraint on the total number of selected items.
The mainstream workflow of image recognition applications is first training one global model on the cloud for a wide range of classes and then serving numerous clients, each with heterogeneous images from a small subset of classes to be recognized. From the cloud-client discrepancies on the range of image classes, the recognition model is desired to have strong adaptiveness, intuitively by concentrating the focus on each individual client's local dynamic class subset, while incurring negligible overhead. In this work, we propose to plug a new intra-client and inter-image attention (ICIIA) module into existing backbone recognition models, requiring only one-time cloud-based training to be client-adaptive. In particular, given a target image from a certain client, ICIIA introduces multi-head self-attention to retrieve relevant images from the client's historical unlabeled images, thereby calibrating the focus and the recognition result. Further considering that ICIIA's overhead is dominated by linear projection, we propose partitioned linear projection with feature shuffling for replacement and allow increasing the number of partitions to dramatically improve efficiency without scarifying too much accuracy. We finally evaluate ICIIA using 3 different recognition tasks with 9 backbone models over 5 representative datasets. Extensive evaluation results demonstrate the effectiveness and efficiency of ICIIA. Specifically, for ImageNet-1K with the backbone models of MobileNetV3-L and Swin-B, ICIIA can improve the testing accuracy to 83.37% (+8.11%) and 88.86% (+5.28%), while adding only 1.62% and 0.02% of FLOPs, respectively.
In this paper, we study the adaptive submodular cover problem under the worst-case setting. This problem generalizes many previously studied problems, namely, the pool-based active learning and the stochastic submodular set cover. The input of our problem is a set of items (e.g., medical tests) and each item has a random state (e.g., the outcome of a medical test), whose realization is initially unknown. One must select an item at a fixed cost in order to observe its realization. There is an utility function which is defined over items and their states. Our goal is to sequentially select a group of items to achieve a ``goal value'' while minimizing the maximum cost across realizations (a.k.a. worst-case cost). To facilitate our study, we introduce a broad class of stochastic functions, called \emph{worst-case submodular function}. Assume the utility function is worst-case submodular, we develop a tight $(\log (Q/\eta)+1)$-approximation policy, where $Q$ is the ``goal value'' and $\eta$ is the minimum gap between $Q$ and any attainable utility value $\hat{Q}<Q$. We also study a worst-case maximum-coverage problem, whose goal is to select a group of items to maximize its worst-case utility subject to a budget constraint. This is a flipped problem of the minimum-cost-cover problem, and to solve this problem, we develop a $(1-1/e)/2$-approximation solution.
Theoretical studies on evolutionary algorithms have developed vigorously in recent years. Many such algorithms have theoretical guarantees in both running time and approximation ratio. Some approximation mechanism seems to be inherently embedded in many evolutionary algorithms. In this paper, we identify such a relation by proposing a unified analysis framework for a generalized simple multi-objective evolutionary algorithm (GSEMO), and apply it on a minimum weight general cover problem. For a wide range of problems (including the the minimum submodular cover problem in which the submodular function is real-valued, and the minimum connected dominating set problem for which the potential function is non-submodular), GSEMO yields asymptotically tight approximation ratios in expected polynomial time.