Selective Attention: Enhancing Transformer through Principled Context Control

The attention mechanism within the transformer architecture enables the model to weigh and combine tokens based on their relevance to the query. While self-attention has enjoyed major success, it notably treats all queries $q$ in the same way by applying the mapping $V^\top\text{softmax}(Kq)$, where $V,K$ are the value and key embeddings respectively. In this work, we argue that this uniform treatment hinders the ability to control contextual sparsity and relevance. As a solution, we introduce the $\textit{Selective Self-Attention}$ (SSA) layer that augments the softmax nonlinearity with a principled temperature scaling strategy. By controlling temperature, SSA adapts the contextual sparsity of the attention map to the query embedding and its position in the context window. Through theory and experiments, we demonstrate that this alleviates attention dilution, aids the optimization process, and enhances the model's ability to control softmax spikiness of individual queries. We also incorporate temperature scaling for value embeddings and show that it boosts the model's ability to suppress irrelevant/noisy tokens. Notably, SSA is a lightweight method which introduces less than 0.5% new parameters through a weight-sharing strategy and can be fine-tuned on existing LLMs. Extensive empirical evaluations demonstrate that SSA-equipped models achieve a noticeable and consistent accuracy improvement on language modeling benchmarks.

Towards Data Valuation via Asymmetric Data Shapley

As data emerges as a vital driver of technological and economic advancements, a key challenge is accurately quantifying its value in algorithmic decision-making. The Shapley value, a well-established concept from cooperative game theory, has been widely adopted to assess the contribution of individual data sources in supervised machine learning. However, its symmetry axiom assumes all players in the cooperative game are homogeneous, which overlooks the complex structures and dependencies present in real-world datasets. To address this limitation, we extend the traditional data Shapley framework to asymmetric data Shapley, making it flexible enough to incorporate inherent structures within the datasets for structure-aware data valuation. We also introduce an efficient $k$-nearest neighbor-based algorithm for its exact computation. We demonstrate the practical applicability of our framework across various machine learning tasks and data market contexts. The code is available at: https://github.com/xzheng01/Asymmetric-Data-Shapley.

Uncertainty Quantification of Data Shapley via Statistical Inference

As data plays an increasingly pivotal role in decision-making, the emergence of data markets underscores the growing importance of data valuation. Within the machine learning landscape, Data Shapley stands out as a widely embraced method for data valuation. However, a limitation of Data Shapley is its assumption of a fixed dataset, contrasting with the dynamic nature of real-world applications where data constantly evolves and expands. This paper establishes the relationship between Data Shapley and infinite-order U-statistics and addresses this limitation by quantifying the uncertainty of Data Shapley with changes in data distribution from the perspective of U-statistics. We make statistical inferences on data valuation to obtain confidence intervals for the estimations. We construct two different algorithms to estimate this uncertainty and provide recommendations for their applicable situations. We also conduct a series of experiments on various datasets to verify asymptotic normality and propose a practical trading scenario enabled by this method.

Double Variance Reduction: A Smoothing Trick for Composite Optimization Problems without First-Order Gradient

Variance reduction techniques are designed to decrease the sampling variance, thereby accelerating convergence rates of first-order (FO) and zeroth-order (ZO) optimization methods. However, in composite optimization problems, ZO methods encounter an additional variance called the coordinate-wise variance, which stems from the random gradient estimation. To reduce this variance, prior works require estimating all partial derivatives, essentially approximating FO information. This approach demands O(d) function evaluations (d is the dimension size), which incurs substantial computational costs and is prohibitive in high-dimensional scenarios. This paper proposes the Zeroth-order Proximal Double Variance Reduction (ZPDVR) method, which utilizes the averaging trick to reduce both sampling and coordinate-wise variances. Compared to prior methods, ZPDVR relies solely on random gradient estimates, calls the stochastic zeroth-order oracle (SZO) in expectation $\mathcal{O}(1)$ times per iteration, and achieves the optimal $\mathcal{O}(d(n + \kappa)\log (\frac{1}{\epsilon}))$ SZO query complexity in the strongly convex and smooth setting, where $\kappa$ represents the condition number and $\epsilon$ is the desired accuracy. Empirical results validate ZPDVR's linear convergence and demonstrate its superior performance over other related methods.

FLASH: Federated Learning Across Simultaneous Heterogeneities

The key premise of federated learning (FL) is to train ML models across a diverse set of data-owners (clients), without exchanging local data. An overarching challenge to this date is client heterogeneity, which may arise not only from variations in data distribution, but also in data quality, as well as compute/communication latency. An integrated view of these diverse and concurrent sources of heterogeneity is critical; for instance, low-latency clients may have poor data quality, and vice versa. In this work, we propose FLASH(Federated Learning Across Simultaneous Heterogeneities), a lightweight and flexible client selection algorithm that outperforms state-of-the-art FL frameworks under extensive sources of heterogeneity, by trading-off the statistical information associated with the client's data quality, data distribution, and latency. FLASH is the first method, to our knowledge, for handling all these heterogeneities in a unified manner. To do so, FLASH models the learning dynamics through contextual multi-armed bandits (CMAB) and dynamically selects the most promising clients. Through extensive experiments, we demonstrate that FLASH achieves substantial and consistent improvements over state-of-the-art baselines -- as much as 10% in absolute accuracy -- thanks to its unified approach. Importantly, FLASH also outperforms federated aggregation methods that are designed to handle highly heterogeneous settings and even enjoys a performance boost when integrated with them.

Plug-and-Play Transformer Modules for Test-Time Adaptation

Parameter-efficient tuning (PET) methods such as LoRA, Adapter, and Visual Prompt Tuning (VPT) have found success in enabling adaptation to new domains by tuning small modules within a transformer model. However, the number of domains encountered during test time can be very large, and the data is usually unlabeled. Thus, adaptation to new domains is challenging; it is also impractical to generate customized tuned modules for each such domain. Toward addressing these challenges, this work introduces PLUTO: a Plug-and-pLay modUlar Test-time domain adaptatiOn strategy. We pre-train a large set of modules, each specialized for different source domains, effectively creating a ``module store''. Given a target domain with few-shot unlabeled data, we introduce an unsupervised test-time adaptation (TTA) method to (1) select a sparse subset of relevant modules from this store and (2) create a weighted combination of selected modules without tuning their weights. This plug-and-play nature enables us to harness multiple most-relevant source domains in a single inference call. Comprehensive evaluations demonstrate that PLUTO uniformly outperforms alternative TTA methods and that selecting $\leq$5 modules suffice to extract most of the benefit. At a high level, our method equips pre-trained transformers with the capability to dynamically adapt to new domains, motivating a new paradigm for efficient and scalable domain adaptation.

PPFL: A Personalized Federated Learning Framework for Heterogeneous Population

Personalization aims to characterize individual preferences and is widely applied across many fields. However, conventional personalized methods operate in a centralized manner and potentially expose the raw data when pooling individual information. In this paper, with privacy considerations, we develop a flexible and interpretable personalized framework within the paradigm of Federated Learning, called PPFL (Population Personalized Federated Learning). By leveraging canonical models to capture fundamental characteristics among the heterogeneous population and employing membership vectors to reveal clients' preferences, it models the heterogeneity as clients' varying preferences for these characteristics and provides substantial insights into client characteristics, which is lacking in existing Personalized Federated Learning (PFL) methods. Furthermore, we explore the relationship between our method and three main branches of PFL methods: multi-task PFL, clustered FL, and decoupling PFL, and demonstrate the advantages of PPFL. To solve PPFL (a non-convex constrained optimization problem), we propose a novel random block coordinate descent algorithm and present the convergence property. We conduct experiments on both pathological and practical datasets, and the results validate the effectiveness of PPFL.

Causal Rule Learning: Enhancing the Understanding of Heterogeneous Treatment Effect via Weighted Causal Rules

Interpretability is a key concern in estimating heterogeneous treatment effects using machine learning methods, especially for healthcare applications where high-stake decisions are often made. Inspired by the Predictive, Descriptive, Relevant framework of interpretability, we propose causal rule learning which finds a refined set of causal rules characterizing potential subgroups to estimate and enhance our understanding of heterogeneous treatment effects. Causal rule learning involves three phases: rule discovery, rule selection, and rule analysis. In the rule discovery phase, we utilize a causal forest to generate a pool of causal rules with corresponding subgroup average treatment effects. The selection phase then employs a D-learning method to select a subset of these rules to deconstruct individual-level treatment effects as a linear combination of the subgroup-level effects. This helps to answer an ignored question by previous literature: what if an individual simultaneously belongs to multiple groups with different average treatment effects? The rule analysis phase outlines a detailed procedure to further analyze each rule in the subset from multiple perspectives, revealing the most promising rules for further validation. The rules themselves, their corresponding subgroup treatment effects, and their weights in the linear combination give us more insights into heterogeneous treatment effects. Simulation and real-world data analysis demonstrate the superior performance of causal rule learning on the interpretable estimation of heterogeneous treatment effect when the ground truth is complex and the sample size is sufficient.

FedYolo: Augmenting Federated Learning with Pretrained Transformers

The growth and diversity of machine learning applications motivate a rethinking of learning with mobile and edge devices. How can we address diverse client goals and learn with scarce heterogeneous data? While federated learning aims to address these issues, it has challenges hindering a unified solution. Large transformer models have been shown to work across a variety of tasks achieving remarkable few-shot adaptation. This raises the question: Can clients use a single general-purpose model, rather than custom models for each task, while obeying device and network constraints? In this work, we investigate pretrained transformers (PTF) to achieve these on-device learning goals and thoroughly explore the roles of model size and modularity, where the latter refers to adaptation through modules such as prompts or adapters. Focusing on federated learning, we demonstrate that: (1) Larger scale shrinks the accuracy gaps between alternative approaches and improves heterogeneity robustness. Scale allows clients to run more local SGD epochs which can significantly reduce the number of communication rounds. At the extreme, clients can achieve respectable accuracy locally highlighting the potential of fully-local learning. (2) Modularity, by design, enables $>$100$\times$ less communication in bits. Surprisingly, it also boosts the generalization capability of local adaptation methods and the robustness of smaller PTFs. Finally, it enables clients to solve multiple unrelated tasks simultaneously using a single PTF, whereas full updates are prone to catastrophic forgetting. These insights on scale and modularity motivate a new federated learning approach we call "You Only Load Once" (FedYolo): The clients load a full PTF model once and all future updates are accomplished through communication-efficient modules with limited catastrophic-forgetting, where each task is assigned to its own module.

Privacy-Preserving Community Detection for Locally Distributed Multiple Networks

Modern multi-layer networks are commonly stored and analyzed in a local and distributed fashion because of the privacy, ownership, and communication costs. The literature on the model-based statistical methods for community detection based on these data is still limited. This paper proposes a new method for consensus community detection and estimation in a multi-layer stochastic block model using locally stored and computed network data with privacy protection. A novel algorithm named privacy-preserving Distributed Spectral Clustering (ppDSC) is developed. To preserve the edges' privacy, we adopt the randomized response (RR) mechanism to perturb the network edges, which satisfies the strong notion of differential privacy. The ppDSC algorithm is performed on the squared RR-perturbed adjacency matrices to prevent possible cancellation of communities among different layers. To remove the bias incurred by RR and the squared network matrices, we develop a two-step bias-adjustment procedure. Then we perform eigen-decomposition on the debiased matrices, aggregation of the local eigenvectors using an orthogonal Procrustes transformation, and k-means clustering. We provide theoretical analysis on the statistical errors of ppDSC in terms of eigen-vector estimation. In addition, the blessings and curses of network heterogeneity are well-explained by our bounds.