arXiv subjects

Igor Markov

Publications and source records attributed to Igor Markov.

6 recordsLinked to original sources

arXiv2023

RusTitW: Russian Language Text Dataset for Visual Text in-the-Wild Recognition

Information surrounds people in modern life. Text is a very efficient type of information that people use for communication for centuries. However, automated text-in-the-wild recognition remains a challenging problem. The major limitation for a DL system is the lack of training data. For the competitive performance, training set must contain many samples that replicate the real-world cases. While there are many high-quality datasets for English text recognition; there are no available datasets for Russian language. In this paper, we present a large-scale human-labeled dataset for Russian text recognition in-the-wild. We also publish a synthetic dataset and code to reproduce the generation process

cs.CV↗

arXiv2023

Enhancing quantum computer performance via symmetrization

Large quantum computers promise to solve some critical problems not solvable otherwise. However, modern quantum technologies suffer various imperfections such as control errors and qubit decoherence, inhibiting their potential utility. The overheads of quantum error correction are too great for near-term quantum computers, whereas error-mitigation strategies that address specific device imperfections may lose relevance as devices improve. To enhance the performance of quantum computers with high-quality qubits, we introduce a strategy based on symmetrization and nonlinear aggregation. On a commercial trapped-ion quantum computer, it improves performance of multiple practical algorithms by 100x with no qubit or gate overhead.

quant-ph↗

arXiv2022

Federated Calibration and Evaluation of Binary Classifiers

We address two major obstacles to practical use of supervised classifiers on distributed private data. Whether a classifier was trained by a federation of cooperating clients or trained centrally out of distribution, (1) the output scores must be calibrated, and (2) performance metrics must be evaluated -- all without assembling labels in one place. In particular, we show how to perform calibration and compute precision, recall, accuracy and ROC-AUC in the federated setting under three privacy models (i) secure aggregation, (ii) distributed differential privacy, (iii) local differential privacy. Our theorems and experiments clarify tradeoffs between privacy, accuracy, and data efficiency. They also help decide whether a given application has sufficient data to support federated calibration and evaluation.

cs.CR↗

arXiv2022

Data-Driven Mitigation of Adversarial Text Perturbation

Social networks have become an indispensable part of our lives, with billions of people producing ever-increasing amounts of text. At such scales, content policies and their enforcement become paramount. To automate moderation, questionable content is detected by Natural Language Processing (NLP) classifiers. However, high-performance classifiers are hampered by misspellings and adversarial text perturbations. In this paper, we classify intentional and unintentional adversarial text perturbation into ten types and propose a deobfuscation pipeline to make NLP models robust to such perturbations. We propose Continuous Word2Vec (CW2V), our data-driven method to learn word embeddings that ensures that perturbations of words have embeddings similar to those of the original words. We show that CW2V embeddings are generally more robust to text perturbations than embeddings based on character ngrams. Our robust classification pipeline combines deobfuscation and classification, using proposed defense methods and word embeddings to classify whether Facebook posts are requesting engagement such as likes. Our pipeline results in engagement bait classification that goes from 0.70 to 0.67 AUC with adversarial text perturbation, while character ngram-based word embedding methods result in downstream classification that goes from 0.76 to 0.64.

cs.CL↗

arXiv2021

Picasso: Model-free Feature Visualization

Today, Machine Learning (ML) applications can have access to tens of thousands of features. With such feature sets, efficiently browsing and curating subsets of most relevant features is a challenge. In this paper, we present a novel approach to visualize up to several thousands of features in a single image. The image not only shows information on individual features, but also expresses feature interactions via the relative positioning of features.

cs.HC↗

arXiv2007

Constant-degree graph expansions that preserve the treewidth

Many hard algorithmic problems dealing with graphs, circuits, formulas and constraints admit polynomial-time upper bounds if the underlying graph has small treewidth. The same problems often encourage reducing the maximal degree of vertices to simplify theoretical arguments or address practical concerns. Such degree reduction can be performed through a sequence of splittings of vertices, resulting in an _expansion_ of the original graph. We observe that the treewidth of a graph may increase dramatically if the splittings are not performed carefully. In this context we address the following natural question: is it possible to reduce the maximum degree to a constant without substantially increasing the treewidth? Our work answers the above question affirmatively. We prove that any simple undirected graph G=(V, E) admits an expansion G'=(V', E') with the maximum degree <= 3 and treewidth(G') <= treewidth(G)+1. Furthermore, such an expansion will have no more than 2|E|+|V| vertices and 3|E| edges; it can be computed efficiently from a tree-decomposition of G. We also construct a family of examples for which the increase by 1 in treewidth cannot be avoided.

cs.DM↗