Chen Qian’s research while affiliated with University of California, Santa Cruz and other places

Federated fine-tuning (FedFT) has been proposed to fine-tune the pre-trained language models in a distributed manner. However, there are two critical challenges for efficient FedFT in practical applications, i.e., resource constraints and system heterogeneity. Existing works rely on parameter-efficient fine-tuning methods, e.g., low-rank adaptation (LoRA), but with major limitations. Herein, based on the inherent characteristics of FedFT, we observe that LoRA layers with higher ranks added close to the output help to save resource consumption while achieving comparable fine-tuning performance. Then we propose a novel LoRA-based FedFT framework, termed LEGEND, which faces the difficulty of determining the number of LoRA layers (called, LoRA depth) and the rank of each LoRA layer (called, rank distribution). We analyze the coupled relationship between LoRA depth and rank distribution, and design an efficient LoRA configuration algorithm for heterogeneous devices, thereby promoting fine-tuning efficiency. Extensive experiments are conducted on a physical platform with 80 commercial devices. The results show that LEGEND can achieve a speedup of 1.5-2.8$\times$ and save communication costs by about 42.3% when achieving the target accuracy, compared to the advanced solutions.

Instruction tuning is critical for adapting large language models (LLMs) to downstream tasks, and recent studies have demonstrated that small amounts of human-curated data can outperform larger datasets, challenging traditional data scaling laws. While LLM-based data quality rating systems offer a cost-effective alternative to human annotation, they often suffer from inaccuracies and biases, even in powerful models like GPT-4. In this work, we introduce DS2, a Diversity-aware Score curation method for Data Selection. By systematically modeling error patterns through a score transition matrix, DS2 corrects LLM-based scores and promotes diversity in the selected data samples. Our approach shows that a curated subset (just 3.3% of the original dataset) outperforms full-scale datasets (300k samples) across various machine-alignment benchmarks, and matches or surpasses human-aligned datasets such as LIMA with the same sample size (1k samples). These findings challenge conventional data scaling assumptions, highlighting that redundant, low-quality samples can degrade performance and reaffirming that "more can be less."

Federated semi-supervised learning (FSSL) has been proposed to address the insufficient labeled data problem by training models with pseudo-labeling. In previous FSSL systems, a single global model is always trained without an equivalent generalization ability for the clients under the non-IID setting. Accordingly, model personalization methods have been proposed to overcome this problem. Intuitively, seeking labeling assistance from other clients with similar data distribution, i.e. , model migration, can effectively improve the personalization on the clients with scarce labeled data. However, previous works require to migrate a pre-fixed number of models among the clients, causing unnecessary resource waste and accuracy degradation due to resource heterogeneity. Considering that the number of model migrations and the quality of pseudo-labels have a significant impact on the training performance ( e.g. , efficiency and accuracy), we propose a novel personalized FSSL system, called Ferrari, to boost the efficiency of pseudo-labeling and training accuracy through adaptive model migrations among the clients. Specifically, Ferrari first generates the similarity-based ranking using a Gaussian KD-Tree, considering the varied data distributions among the clients. Combined with the ranking and clients' heterogeneous resource constraints, Ferrari then adaptively determines the proper model migration policy and confidence thresholds for high-quality pseudo-labeling and personalized training for clients. Extensive experiments on a physical platform show that Ferrari provides a 1.2 $\sim 5.5\times$ speedup without sacrificing model accuracy, compared to existing methods.

Decentralized federated learning (DFL) uses peer-to-peer communication to avoid the single point of failure problem in federated learning and has been considered an attractive solution for machine learning tasks on distributed devices. We provide the first solution to a fundamental network problem of DFL: what overlay network should DFL use to achieve fast training of highly accurate models, low communication, and decentralized construction and maintenance? Overlay topologies of DFL have been investigated, but no existing DFL topology includes decentralized protocols for network construction and topology maintenance. Without these protocols, DFL cannot run in practice. This work presents an overlay network, called FedLay, which provides fast training and low communication cost for practical DFL. FedLay is the first solution for constructing near-random regular topologies in a decentralized manner and maintaining the topologies under node joins and failures. Experiments based on prototype implementation and simulations show that FedLay achieves the fastest model convergence and highest accuracy on real datasets compared to existing DFL solutions while incurring small communication costs and being resilient to node joins and failures.

To provide a flexible tradeoff between inference accuracy and resource requirement at runtime, the slimmable neural network (SNN), a single network executable at different widths with the same deploying and management cost as that of a single model, has been proposed. However, how to effectively train SNN among massive devices in edge computing without revealing their local data remains an open problem. To this end, we leverage a novel distributed machine learning paradigm, i.e. , federated learning, to realize effective on-device SNN training. As current FL schemes often train only one model with fixed architecture, and the existing SNN training algorithm is resource-intensive, integrating FL and SNN is non-trivial. Furthermore, two intrinsic features in edge computing, i.e. , data and system heterogeneity, exacerbate the difficulty. Motivated by this, we redesign the model distribution, local training, and model aggregation phases in traditional FL, and propose FedSNN, a framework that ensures all widths in SNN can obtain high accuracy with less resource consumption. Specifically, for devices with heterogeneous training capacities and data distributions, the parameter server will distribute each of them with one proper width for adaptive local training guided by their uploaded model features, and their trained models will be weighted-averaged using the proposed multi-width SNN aggregation to improve their statistical utility. Extensive experiments on a distributed testbed show that FedSNN improves the model accuracy by about 2.18%-8.1%, and accelerates training by about 1.31 $\times$ -6.84 $\times$ , compared with existing solutions.

Federated learning (FL) is an efficient distributed learning method that facilitates collaborative model training among multiple edge devices (or clients). However, current research always assumes that clients have access to ground-truth data for training, which is unrealistic in practice because of a lack of expertise. Semi-supervised federated learning (SSFL) has been proposed in many existing works to address this problem, which always adopts a fixed model architecture for training, bringing two main problems with varying amounts of pseudo-labeled data. First, the shallow model cannot have the capability to fit the increasing pseudo-labeled data, leading to poor training performance. Second, the large model suffers from an overfitting problem when exploiting a few labeled data samples in SSFL, and also requires tremendous resource ( e.g. , computation and communication) costs. To tackle these problems, we propose a novel framework, called STAR , which adopts progressive training to enhance model training in SSFL. Specifically, STAR gradually increases the model depth through adding the sub-module ( e.g. , one or several layers) from a shallow model, and performs pseudo-labeling for unlabeled data with a specialized confidence threshold simultaneously. Then, we propose an efficient algorithm to determine the appropriate model depth for each client with varied resource budgets and the proper confidence threshold for pseudo-labeling in SSFL. Our proposed framework STAR innovatively applies progressive training to SSFL, which significantly contributes to the advancement of the FL field. STAR has been evaluated through extensive experiments, and the results demonstrate its high effectiveness. For instance, STAR can reduce the bandwidth consumption by about 40%, and achieve an average accuracy improvement of around 9.8% compared with the baselines, on CIFAR10. Besides, STAR achieves about 2.2× speedup compared to the baselines on ImageNet100.