Patekote į pirmąjį lietuvišką domeną. 
2026 m. gegužės 18 d., 13 val.
Vilnius, Akademijos g. 4, 203 kab.
Nuotoliniu būdu „MS Teams“ aplinkoje (https://bit.ly/DMSTI_2026-05-18)
Justina Ramonaitė
(vadovė dr. Gražina Korvel)
„Giliuoju mokymusi pagrįstas šnekos signalo gerinimas“
Santrauka: Seminaro metu bus pristatyti vykdomi tyrimai su generatyviniu adversariniu tinklu, kuris skirtas triukšmo šalinimui iš šnekos signalo. Pirmasis eksperimentas skirtas palyginti, kaip skiriasi gaunami rezultatai, kai skiriasi kalbėtojo akcentas. Antras eksperimentas skirtas patikrinimui, ar nuostolių funkcijos papildymas SI-SNR (scale-invariant signal-to-noise ratio) informacija padidina modelio efektyvumą
Ieva Ramašauskienė
(vadovas dr. Andrius Čiginas)
„Netikimybinės imties integravimas naudojant kombinuotus įverčius“
Santrauka: As survey response rates decline and alternative data sources become increasingly available, non-probability samples offer a faster and more cost-effective option for statistical inference. However, their use requires careful methodological treatment because the sample selection mechanism is typically unknown. If ignored, this can result in biased population parameter estimates. Our study develops methods for improving inference from non-probability samples under the Not-Missing-At-Random selection mechanism. We extend propensity score adjustment methodology for cases where complete auxiliary data are available and introduce an analytical variance estimation approach for the inverse probability weighted (IPW) estimator to improve computational efficiency. In addition, we propose and compare two types of composite estimators that combine the IPW estimator based on the non-probability sample with the direct estimator obtained from a pooled non-probability and complementary probability sample. Simulation results based on Lithuanian Census data demonstrate the efficiency and practical value of the proposed techniques.