Coupled training for seq2seq models
Organisatsiooni nimi
TartuNLP
Kokkuvõte
The goal is to develop and test a particular method for training sequence-to-sequence models (for text, speech, image processing and generation) that uses existing pre-trained models from HuggingFace as a guide. As a result, the new trained model shares a compatible vector space with the pre-trained model's encoder/decoder, but can focus on a particular task, represented in the data -- the compatibility avoids catastrophic forgetting and allows to create "extension modules" for existing pre-trained models.
Lõputöö kaitsmise aasta
2024-2025
Juhendaja
Mark Fišel
Suhtlemiskeel(ed)
eesti keel, inglise keel
Nõuded kandideerijale
Familiarity with machine learning and neural networks. Confident Python skills. Some experience with HuggingFace/PyTorch or readiness to learn quickly. Text/image/sound data processing experience or readiness to learn quickly.
Tase
Magister
Kandideerimise kontakt
Nimi
Mark Fišel
Tel
E-mail