YerevaNN

Research

• Machine learning algorithms
  • [2023] Scaling laws for mixed-modal language models. [preprint] (with Meta AI)
  • [2022] Matching map recovery with an unknown number of outliers. Accepted at AISTATS'23 [preprint] (with ENSAE-CREST)
  • [2022] GradSkip: an extension of a local gradient method for distributed optimization that supports variable number of local gradient steps in each communication round. [preprint] (with KAUST)
  • [2021] Characterization of the failure modes of domain generalization algorithms. Published in CVPR'22 (with USC ISI).
  • [2021] WARP: a parameter-efficient method for transfer learning in NLP. Published in ACL'21 (with USC ISI)
  • [2021] Theoretical analysis of the detection of the feature matching map in presence of outliers. Published in Electronic Journal of Statistics (with ENSAE-CREST)
  • [2021] A survey of deep neural networks for semi-supervised image classification. Published in JUCS.
  • [2020] Robust classification under class-dependent domain shift [preprint] (with USC ISI). Presented at ICML 2020 UDL Workshop
  • [2020] A novel robust estimator of the mean of a multivariate Gaussian distribution. Published in Annals of Statistics (with ENSAE-CREST)
  • [2019] T-Corex: a novel method for temporal covariance estimation using information theoretic apparatus [preprint] [code] (with USC ISI)
• Machine learning for biomedical data
  • [2022] BARTSmiles: large-scale generative masked language models for molecular representations . [preprint] [code] (with IMB NAS RA, Toxometris.ai and Meta AI)
  • [2022] Improved molecular representations for property prediction based on VAEs. Published in Journal of Cheminformatics. (with Institute of Molecular Biology NAS RA).
  • [2020] Parasite: a method to align sentences in parallel corpora [code]. Helped to win WMT'20 Biomedical Translation Shared Task for en-ru pairs [paper].
  • [2019] BioRelEx: biological relation extraction dataset [data] [baselines] Published in BioNLP'19 workshop (with USC ISI and Bioinformatics group at IMB NAS RA, as part of the Big Mechanism program)
  • [2018] Clinical time series prediction benchmarks [code] [pretrained models]. Published in Nature Scientific Data (with USC ISI)
• Development of Armenian treebanks
  • Eastern Armenian treebank as part of the Universal Dependencies project [data (52K tokens)] [website] [code]
  • An end-to-end syntax parser for Armenian [code] [demo]
  • Collaboration with Marat Yavrumyan and Anna Danielyan from Yerevan State University. Partially funded by ANSEF and ISTC.
• Student projects
  • [2019] Morpheme-aware word vectors [paper] [code]
  • [2018] A joint POS tagger and lemmatizer [paper] [code]
  • [2018] Reproducing DIIN network for NLI [report] [code]
  • [2017] Reproducing R-NET network for QA [blogpost] [code]