Prior Knowledge for Training Neural Seq2Seq Models
NAVER LABS Europe (NLE) is opening a research internship with the goal of advancing the use of prior knowledge for training neural seq2seq models.
When training data is scarce, as is often the case in practice, standard end-to-end training of sequential models tends to produce mediocre results; however, it is often the case that the developer of such models is aware of structural biases and global characteristics of candidates that could be exploited to better generalize from the available direct observations.
We are looking for a motivated intern to join an ongoing research project addressing this general problem, both in theory and in practice. The experiments will be conducted both on synthetic data, for permitting a fine-grained and flexible exploration of different parameters of the problem, as well as on natural data (e.g. NLG, MT), for validating the usefulness of the techniques.
The successful candidate should be enrolled in a graduate program, at the Master or (preferably) PhD level, with a focus on Deep Learning (knowledge of NLP and RL a plus).
Strong mathematical and programming skills as well as familiarity with one of the major current deep learning toolkits (PyTorch preferred but not compulsory) are a requirement.
Publication of results in major conferences/journals will be strongly encouraged.