Term papers

DEADLINE: term papers (and code, if any) are due on 20/09/2019 Reports must be no longer than 10 pages, 11pt font. There is no minimal length.

Instructions

Two kinds of term papers are possible.

A scientific report on a paper not discussed in the seminar, but related to deep learning (and ideally, to investigating and understanding the mechanics of a deep model). The report should make clear what kind of problems the chosen paper addresses, and the results it achieves. Ideally, it should also establish relevant comparison with other works in the field (typically, works discussed in the seminar). You may also use any additional reference or ressource you like to this aim. A dissertation making a good job at summarizing the paper’s problematics and findings will be enough for validation. To be considered for maximal grade, the report should in addition, carefully discusses how the work relates to relevant litterature and/or raises and discuss relevant questions.
Some implementation/experimental work and a short written report about the process and results. This may include: implementing a model discussed in the seminar, playing with it on a task of your choice or reproducing some of the experiments of a paper, possibly on a simplified dataset. This can also include downloading the code from the paper’s author and running additional experiments

In both cases, chosen topic must be preempively discussed with and validated by Dr. Venant.

Topics’ suggestions

This are only suggestion, you are more than encouraged to propose your own.

Reproduce this experiments of Nangia & Bowman (2018): “Listops: A Diagnostic Dataset for Latent Tree Learning” (using their dataset). And/or try with different algebraic operations (possibly using their code) (Experimental).
Try some of the LSTM simplifications tested in Levy et al. (2018): “Long Short-Term Memory as a Dynamically Computed Element-wise Weighted Sum” on a task of your choice using Tree-LSTMs (Experimental).
Build a toy neural recognizer or parser. Train it using various artificial data generated using context free or non context free grammars (Experimental).
Report on Shi et al. (2018): “On Tree-Based Neural Sentence Modeling”, in the light of the seminar’s discussions (Report).
Report on Tran et al. (2018) “The Importance of Being Recurrent for Modeling Hierarchical Structure”, in the light of the seminar’s discussions (Report).