arXiv subjects

Antonios Deligiannakis

Publications and source records attributed to Antonios Deligiannakis.

5 recordsLinked to original sources

arXiv2026

FDA-Opt: Federated Fine-Tuning via Dynamic Update Schedules

Federated Learning (FL) enables the utilization of vast, previously inaccessible data sources. At the same time, pre-trained Language Models (LMs) have taken the world by storm and for good reason. They exhibit remarkable emergent abilities and are readily adapted to downstream tasks. This opens one of the most exciting frontiers in FL: fine-tuning LMs. Yet, a persistent challenge in FL is the frequent, rigid communication of parameters -- a problem magnified by the sheer size of these contemporary models. The FedOpt family of algorithms has become the go-to approach for FL, relying on fixed but arbitrary intervals for model exchanges. Recently, the FDA algorithm prescribed a dynamic approach by monitoring the training progress. However, it introduced a hard-to-calibrate parameter and imposed a rigid synchronization scheme. In this work, we address these limitations by proposing the FDA-Opt family of algorithms -- a unified generalization of both FDA and FedOpt. Our experimental evaluation focuses on fine-tuning LMs on downstream NLP tasks and demonstrates that FDA-Opt outperforms FedOpt even when it is configured with hyper-parameters specifically optimized for the latter. In other words, we show that FDA-Opt is a practical, drop-in replacement for FedOpt in modern FL libraries and systems: it requires no additional configuration and delivers superior performance out of the box.

cs.LG↗

arXiv2024

Communication-Efficient Distributed Deep Learning via Federated Dynamic Averaging

The ever-growing volume and decentralized nature of data, coupled with the need to harness it and extract knowledge, have led to the extensive use of distributed deep learning (DDL) techniques for training. These techniques rely on local training performed at distributed nodes using locally collected data, followed by a periodic synchronization process that combines these models to create a unified global model. However, the frequent synchronization of deep learning models, encompassing millions to many billions of parameters, creates a communication bottleneck, severely hindering scalability. Worse yet, DDL algorithms typically waste valuable bandwidth and render themselves less practical in bandwidth-constrained federated settings by relying on overly simplistic, periodic, and rigid synchronization schedules. These inefficiencies make the training process increasingly impractical as they demand excessive time for data communication. To address these shortcomings, we propose Federated Dynamic Averaging (FDA), a communication-efficient DDL strategy that dynamically triggers synchronization based on the value of the model variance. In essence, the costly synchronization step is triggered only if the local models -- initialized from a common global model after each synchronization -- have significantly diverged. This decision is facilitated by the transmission of a small local state from each distributed node. Through extensive experiments across a wide range of learning tasks we demonstrate that FDA reduces communication cost by orders of magnitude, compared to both traditional and cutting-edge communication-efficient algorithms. Additionally, we show that FDA maintains robust performance across diverse data heterogeneity settings.

cs.LG↗

arXiv2020

A Synopses Data Engine for Interactive Extreme-Scale Analytics

In this work, we detail the design and structure of a Synopses Data Engine (SDE) which combines the virtues of parallel processing and stream summarization towards delivering interactive analytics at extreme scale. Our SDE is built on top of Apache Flink and implements a synopsis-as-a-service paradigm. In that it achieves (a) concurrently maintaining thousands of synopses of various types for thousands of streams on demand, (b) reusing maintained synopses among various concurrent workflows, (c) providing data summarization facilities even for cross-(Big Data) platform workflows, (d) pluggability of new synopses on-the-fly, (e) increased potential for workflow execution optimization. The proposed SDE is useful for interactive analytics at extreme scales because it enables (i) enhanced horizontal scalability, i.e., not only scaling out the computation to a number of processing units available in a computer cluster, but also harnessing the processing load assigned to each by operating on carefully-crafted data summaries, (ii) vertical scalability, i.e., scaling the computation to very high numbers of processed streams and (iii) federated scalability i.e., scaling the computation beyond single clusters and clouds by controlling the communication required to answer global queries posed over a number of potentially geo-dispersed clusters.

cs.DB↗

arXiv2012

Sketch-based Querying of Distributed Sliding-Window Data Streams

While traditional data-management systems focus on evaluating single, ad-hoc queries over static data sets in a centralized setting, several emerging applications require (possibly, continuous) answers to queries on dynamic data that is widely distributed and constantly updated. Furthermore, such query answers often need to discount data that is "stale", and operate solely on a sliding window of recent data arrivals (e.g., data updates occurring over the last 24 hours). Such distributed data streaming applications mandate novel algorithmic solutions that are both time- and space-efficient (to manage high-speed data streams), and also communication-efficient (to deal with physical data distribution). In this paper, we consider the problem of complex query answering over distributed, high-dimensional data streams in the sliding-window model. We introduce a novel sketching technique (termed ECM-sketch) that allows effective summarization of streaming data over both time-based and count-based sliding windows with probabilistic accuracy guarantees. Our sketch structure enables point as well as inner-product queries, and can be employed to address a broad range of problems, such as maintaining frequency statistics, finding heavy hitters, and computing quantiles in the sliding-window model. Focusing on distributed environments, we demonstrate how ECM-sketches of individual, local streams can be composed to generate a (low-error) ECM-sketch summary of the order-preserving aggregation of all streams; furthermore, we show how ECM-sketches can be exploited for continuous monitoring of sliding-window queries over distributed streams. Our extensive experimental study with two real-life data sets validates our theoretical claims and verifies the effectiveness of our techniques. To the best of our knowledge, ours is the first work to address efficient, guaranteed-error complex query answ...[truncated].

cs.DB↗

arXiv2011

Discovering Attractive Products based on Influence Sets

Skyline queries have been widely used as a practical tool for multi-criteria decision analysis and for applications involving preference queries. For example, in a typical online retail application, skyline queries can help customers select the most interesting, among a pool of available, products. Recently, reverse skyline queries have been proposed, highlighting the manufacturer's perspective, i.e. how to determine the expected buyers of a given product. In this work we develop novel algorithms for two important classes of queries involving customer preferences. We first propose a novel algorithm, termed as RSA, for answering reverse skyline queries. We then introduce a new type of queries, namely the k-Most Attractive Candidates k-MAC query. In this type of queries, given a set of existing product specifications P, a set of customer preferences C and a set of new candidate products Q, the k-MAC query returns the set of k candidate products from Q that jointly maximizes the total number of expected buyers, measured as the cardinality of the union of individual reverse skyline sets (i.e., influence sets). Applying existing approaches to solve this problem would require calculating the reverse skyline set for each candidate, which is prohibitively expensive for large data sets. We, thus, propose a batched algorithm for this problem and compare its performance against a branch-and-bound variant that we devise. Both of these algorithms use in their core variants of our RSA algorithm. Our experimental study using both synthetic and real data sets demonstrates that our proposed algorithms outperform existing, or naive solutions to our studied classes of queries.

cs.DB↗