arXiv subjects

Nikita Ivkin

Publications and source records attributed to Nikita Ivkin.

13 recordsLinked to original sources

arXiv2026

Federated Learning Clients Clustering with Adaptation to Data Drifts

Federated Learning (FL) trains deep models across edge devices without centralizing raw data, preserving user privacy. However, client heterogeneity slows down convergence and limits global model accuracy. Clustered FL (CFL) mitigates this by grouping clients with similar representations and training a separate model for each cluster. In practice, client data evolves over time, a phenomenon we refer to as data drift, which breaks cluster homogeneity and degrades performance. Data drift can take different forms depending on whether changes occur in the output values, the input features, or the relationship between them. We propose FIELDING, a CFL framework for handling diverse types of data drift with low overhead. FIELDING detects drift at individual clients and performs selective re-clustering to balance cluster quality and model performance, while remaining robust to malicious clients and varying levels of heterogeneity. Experiments show that FIELDING improves final model accuracy by 1.9-5.9% and achieves target accuracy 1.16x-2.23x faster than existing state-of-the-art CFL methods.

cs.LG↗

arXiv2026

XShare: Collaborative in-Batch Expert Sharing for Faster MoE Inference

Mixture-of-Experts (MoE) architectures are increasingly used to efficiently scale large language models. However, in production inference, request batching and speculative decoding significantly amplify expert activation, eroding these efficiency benefits. We address this issue by modeling batch-aware expert selection as a modular optimization problem and designing efficient greedy algorithms for different deployment settings. The proposed method, namely XShare, requires no retraining and dynamically adapts to each batch by maximizing the total gating score of selected experts. It reduces expert activation by up to 30% under standard batching, cuts peak GPU load by up to 3x in expert-parallel deployments, and achieves up to 14% throughput gains in speculative decoding via hierarchical, correlation-aware expert selection even if requests in a batch drawn from heterogeneous datasets.

cs.LG↗

arXiv2020

Amazon SageMaker Autopilot: a white box AutoML solution at scale

AutoML systems provide a black-box solution to machine learning problems by selecting the right way of processing features, choosing an algorithm and tuning the hyperparameters of the entire pipeline. Although these systems perform well on many datasets, there is still a non-negligible number of datasets for which the one-shot solution produced by each particular system would provide sub-par performance. In this paper, we present Amazon SageMaker Autopilot: a fully managed system providing an automated ML solution that can be modified when needed. Given a tabular dataset and the target column name, Autopilot identifies the problem type, analyzes the data and produces a diverse set of complete ML pipelines including feature preprocessing and ML algorithms, which are tuned to generate a leaderboard of candidate models. In the scenario where the performance is not satisfactory, a data scientist is able to view and edit the proposed ML pipelines in order to infuse their expertise and business knowledge without having to revert to a fully manual solution. This paper describes the different components of Autopilot, emphasizing the infrastructure choices that allow scalability, high quality models, editable ML pipelines, consumption of artifacts of offline meta-learning, and a convenient integration with the entire SageMaker suite allowing these trained models to be used in a production setting.

cs.LG↗

arXiv2020

Sketch and Scale: Geo-distributed tSNE and UMAP

Running machine learning analytics over geographically distributed datasets is a rapidly arising problem in the world of data management policies ensuring privacy and data security. Visualizing high dimensional data using tools such as t-distributed Stochastic Neighbor Embedding (tSNE) and Uniform Manifold Approximation and Projection (UMAP) became common practice for data scientists. Both tools scale poorly in time and memory. While recent optimizations showed successful handling of 10,000 data points, scaling beyond million points is still challenging. We introduce a novel framework: Sketch and Scale (SnS). It leverages a Count Sketch data structure to compress the data on the edge nodes, aggregates the reduced size sketches on the master node, and runs vanilla tSNE or UMAP on the summary, representing the densest areas, extracted from the aggregated sketch. We show this technique to be fully parallel, scale linearly in time, logarithmically in memory, and communication, making it possible to analyze datasets with many millions, potentially billions of data points, spread across several data centers around the globe. We demonstrate the power of our method on two mid-size datasets: cancer data with 52 million 35-band pixels from multiple images of tumor biopsies; and astrophysics data of 100 million stars with multi-color photometry from the Sloan Digital Sky Survey (SDSS).

cs.DC↗

arXiv2020

FetchSGD: Communication-Efficient Federated Learning with Sketching

Existing approaches to federated learning suffer from a communication bottleneck as well as convergence issues due to sparse client participation. In this paper we introduce a novel algorithm, called FetchSGD, to overcome these challenges. FetchSGD compresses model updates using a Count Sketch, and then takes advantage of the mergeability of sketches to combine model updates from many workers. A key insight in the design of FetchSGD is that, because the Count Sketch is linear, momentum and error accumulation can both be carried out within the sketch. This allows the algorithm to move momentum and error accumulation from clients to the central aggregator, overcoming the challenges of sparse client participation while still achieving high compression rates and good convergence. We prove that FetchSGD has favorable convergence guarantees, and we demonstrate its empirical effectiveness by training two residual networks and a transformer model.

cs.LG↗

arXiv2020

Practical and sample efficient zero-shot HPO

Zero-shot hyperparameter optimization (HPO) is a simple yet effective use of transfer learning for constructing a small list of hyperparameter (HP) configurations that complement each other. That is to say, for any given dataset, at least one of them is expected to perform well. Current techniques for obtaining this list are computationally expensive as they rely on running training jobs on a diverse collection of datasets and a large collection of randomly drawn HPs. This cost is especially problematic in environments where the space of HPs is regularly changing due to new algorithm versions, or changing architectures of deep networks. We provide an overview of available approaches and introduce two novel techniques to handle the problem. The first is based on a surrogate model and adaptively chooses pairs of dataset, configuration to query. The second, for settings where finding, tuning and testing a surrogate model is problematic, is a multi-fidelity technique combining HyperBand with submodular optimization. We benchmark our methods experimentally on five tasks (XGBoost, LightGBM, CatBoost, MLP and AutoML) and show significant improvement in accuracy compared to standard zero-shot HPO with the same training budget. In addition to contributing new algorithms, we provide an extensive study of the zero-shot HPO technique resulting in (1) default hyper-parameters for popular algorithms that would benefit the community using them, (2) massive lookup tables to further the research of hyper-parameter tuning.

stat.ML↗

arXiv2020

Communication-efficient distributed SGD with Sketching

Large-scale distributed training of neural networks is often limited by network bandwidth, wherein the communication time overwhelms the local computation time. Motivated by the success of sketching methods in sub-linear/streaming algorithms, we introduce Sketched SGD, an algorithm for carrying out distributed SGD by communicating sketches instead of full gradients. We show that Sketched SGD has favorable convergence rates on several classes of functions. When considering all communication -- both of gradients and of updated model weights -- Sketched SGD reduces the amount of communication required compared to other gradient compression methods from $\mathcal{O}(d)$ or $\mathcal{O}(W)$ to $\mathcal{O}(\log d)$, where $d$ is the number of model parameters and $W$ is the number of workers participating in training. We run experiments on a transformer model, an LSTM, and a residual network, demonstrating up to a 40x reduction in total communication cost with no loss in final model performance. We also show experimentally that Sketched SGD scales to at least 256 workers without increasing communication cost or degrading model performance.

cs.LG↗

arXiv2019

Six Dimensional Streaming Algorithm for Cluster Finding in N-Body Simulations

Cosmological N-body simulations are crucial for understanding how the Universe evolves. Studying large-scale distributions of matter in these simulations and comparing them to observations usually involves detecting dense clusters of particles called "halos,'' which are gravitationally bound and expected to form galaxies. However, traditional cluster finders are computationally expensive and use massive amounts of memory. Recent work by Liu et al (Liu et al. (2015)) showed the connection between cluster detection and memory-efficient streaming algorithms and presented a halo finder based on heavy hitter algorithm. Later, Ivkin et al. (Ivkin et al. (2018)) improved the scalability of suggested streaming halo finder with efficient GPU implementation. Both works map particles' positions onto a discrete grid, and therefore lose the rest of the information, such as their velocities. Therefore, two halos travelling through each other are indistinguishable in positional space, while the velocity distribution of those halos can help to identify this process which is worth further studying. In this project we analyze data from the Millennium Simulation Project (Springel et al. (2005)) to motivate the inclusion of the velocity into streaming method we introduce. We then demonstrate a use of suggested method, which allows one to find the same halos as before, while also detecting those which were indistinguishable in prior methods.

astro-ph.GA↗

arXiv2019

Streaming Quantiles Algorithms with Small Space and Update Time

Approximating quantiles and distributions over streaming data has been studied for roughly two decades now. Recently, Karnin, Lang, and Liberty proposed the first asymptotically optimal algorithm for doing so. This manuscript complements their theoretical result by providing a practical variants of their algorithm with improved constants. For a given sketch size, our techniques provably reduce the upper bound on the sketch error by a factor of two. These improvements are verified experimentally. Our modified quantile sketch improves the latency as well by reducing the worst case update time from $O(1/\varepsilon)$ down to $O(\log (1/\varepsilon))$. We also suggest two algorithms for weighted item streams which offer improved asymptotic update times compared to naïve extensions. Finally, we provide a specialized data structure for these sketches which reduces both their memory footprints and update times.

cs.DS↗

arXiv2018

DreamNLP: Novel NLP System for Clinical Report Metadata Extraction using Count Sketch Data Streaming Algorithm: Preliminary Results

Extracting information from electronic health records (EHR) is a challenging task since it requires prior knowledge of the reports and some natural language processing algorithm (NLP). With the growing number of EHR implementations, such knowledge is increasingly challenging to obtain in an efficient manner. We address this challenge by proposing a novel methodology to analyze large sets of EHRs using a modified Count Sketch data streaming algorithm termed DreamNLP. By using DreamNLP, we generate a dictionary of frequently occurring terms or heavy hitters in the EHRs using low computational memory compared to conventional counting approach other NLP programs use. We demonstrate the extraction of the most important breast diagnosis features from the EHRs in a set of patients that underwent breast imaging. Based on the analysis, extraction of these terms would be useful for defining important features for downstream tasks such as machine learning for precision medicine.

cs.LG↗

arXiv2018

Scalable Streaming Tools for Analyzing $N$-body Simulations: Finding Halos and Investigating Excursion Sets in One Pass

Cosmological $N$-body simulations play a vital role in studying models for the evolution of the Universe. To compare to observations and make a scientific inference, statistic analysis on large simulation datasets, e.g., finding halos, obtaining multi-point correlation functions, is crucial. However, traditional in-memory methods for these tasks do not scale to the datasets that are forbiddingly large in modern simulations. Our prior paper proposes memory-efficient streaming algorithms that can find the largest halos in a simulation with up to $10^9$ particles on a small server or desktop. However, this approach fails when directly scaling to larger datasets. This paper presents a robust streaming tool that leverages state-of-the-art techniques on GPU boosting, sampling, and parallel I/O, to significantly improve performance and scalability. Our rigorous analysis of the sketch parameters improves the previous results from finding the centers of the $10^3$ largest halos to $\sim 10^4-10^5$, and reveals the trade-offs between memory, running time and number of halos. Our experiments show that our tool can scale to datasets with up to $\sim 10^{12}$ particles while using less than an hour of running time on a single GPU Nvidia GTX 1080.

astro-ph.IM↗

arXiv2017

BPTree: an $\ell_2$ heavy hitters algorithm using constant memory

The task of finding heavy hitters is one of the best known and well studied problems in the area of data streams. One is given a list $i_1,i_2,\ldots,i_m\in[n]$ and the goal is to identify the items among $[n]$ that appear frequently in the list. In sub-polynomial space, the strongest guarantee available is the $\ell_2$ guarantee, which requires finding all items that occur at least $ε\|f\|_2$ times in the stream, where the vector $f\in\mathbb{R}^n$ is the count histogram of the stream with $i$th coordinate equal to the number of times~$i$ appears $f_i:=\#\{j\in[m]:i_j=i\}$. The first algorithm to achieve the $\ell_2$ guarantee was the CountSketch of [CCF04], which requires $O(ε^{-2}\log n)$ words of memory and $O(\log n)$ update time and is known to be space-optimal if the stream allows for deletions. The recent work of [BCIW16] gave an improved algorithm for insertion-only streams, using only $O(ε^{-2}\logε^{-1}\log\log n)$ words of memory. In this work, we give an algorithm \bptree for $\ell_2$ heavy hitters in insertion-only streams that achieves $O(ε^{-2}\logε^{-1})$ words of memory and $O(\logε^{-1})$ update time, which is the optimal dependence on $n$ and $m$. In addition, we describe an algorithm for tracking $\|f\|_2$ at all times with $O(ε^{-2})$ memory and update time. Our analyses rely on bounding the expected supremum of a Bernoulli process involving Rademachers with limited independence, which we accomplish via a Dudley-like chaining argument that may have applications elsewhere.

cs.DS↗

arXiv2015

Beating CountSketch for Heavy Hitters in Insertion Streams

Given a stream $p_1, \ldots, p_m$ of items from a universe $\mathcal{U}$, which, without loss of generality we identify with the set of integers $\{1, 2, \ldots, n\}$, we consider the problem of returning all $\ell_2$-heavy hitters, i.e., those items $j$ for which $f_j \geq ε\sqrt{F_2}$, where $f_j$ is the number of occurrences of item $j$ in the stream, and $F_2 = \sum_{i \in [n]} f_i^2$. Such a guarantee is considerably stronger than the $\ell_1$-guarantee, which finds those $j$ for which $f_j \geq εm$. In 2002, Charikar, Chen, and Farach-Colton suggested the {\sf CountSketch} data structure, which finds all such $j$ using $Θ(\log^2 n)$ bits of space (for constant $ε> 0$). The only known lower bound is $Ω(\log n)$ bits of space, which comes from the need to specify the identities of the items found. In this paper we show it is possible to achieve $O(\log n \log \log n)$ bits of space for this problem. Our techniques, based on Gaussian processes, lead to a number of other new results for data streams, including (1) The first algorithm for estimating $F_2$ simultaneously at all points in a stream using only $O(\log n\log\log n)$ bits of space, improving a natural union bound and the algorithm of Huang, Tai, and Yi (2014). (2) A way to estimate the $\ell_{\infty}$ norm of a stream up to additive error $ε\sqrt{F_2}$ with $O(\log n\log\log n)$ bits of space, resolving Open Question 3 from the IITK 2006 list for insertion only streams.

cs.DS↗