-
Medical code classification based on free-text clinical notes
(09/2023)
Guillaume Kunsch, Mickael Assaraf, Alexander Belikov
[Thesis]
TL;DR: Thesis on medical code
classification under the supervision of Dr Alexander Belikov.
We use RoBERTa-PM to encode the clinical notes on top of which we create a
custom hierarchical decoder to beat current SOTA on rare codes.
Besides, we hinge the potential of out-of-distribution data in
training and the structure of hyperbolic space for representing
codes hierarchy.
-
Deep RL for multi-agent interaction (03/2023)
Marc Schachtsiek, Guillaume Kunsch
[Code]
TL;DR: A Deep Reinforcement Learning project for
Lux AI
challenge about multi-agent optimal policy for survival in a hostile
environment. This was on open project where we had to design our own
loss function based on multi-objectives in the game, and find a way
to handle the multiple instability due to large learning space. We
applied concepts seen in the course of theoretical RL and deep RL
such as PPO, TRPO or AC2 and optimize for agents behavior against an
opponent. Interestingly, a combination of rule-based + RL-based
behavior performed better both in our case and at the competition
level. A reminder that ML is not always the best answer.
-
Automated essay scoring (03/2023)
Hippolyte Gisserot, Guillaume Kunsch
[Code + Report]
TL;DR: A quick project done during NLP class with Hippolyte
Gisserot. We benchmark several models - LSTM, CNN, Transformers - on
the Kaggle
ASAP
challenge dataset, whose goal is to give a mark for an essay based
on its content. We use historical word-level tokenization methods
such as Word2Vec and GloVe. Interestingly, without any
hyperparameters tuning, CNNs performs on par with the best
Transformer model.
-
Adversarial attacks for image classification (01/2023)
Benjamin Sykes, Guillaume Kunsch, Hippolyte Gisserot
[Code + Report]
TL;DR: In this work, we study the robustness of neural
network classification models to adversarial white-box and black-box
attacks. We implement the Fast Gradient Sign Method (FGSM) and the
Projected Gradient Descent (PGD) attack and evaluate their
performance on CIFAR-10. We also study the impact of adversarial
training on perturbed images and the way to leverage a VAE as a way
of introducing noise.
-
Optimization for ML (12/2022)
Guillaume Kunsch
[Code]
TL;DR: A collection of notebooks implementing various
algorithms whose theoretical performance was studied in the course
Optimization for Machine Learning given by
Gabriel Peyré. Algorithms
implemented: SGD, some Autodiff from scratch, PGD, Proximal
Gradients, Heavy Ball, Nesterov, ...
-
Quantitative study on the relationship between cryptocurrencies
and social media flows (06/2021)
Guillaume Kunsch
[Code + Report]
TL;DR: Thesis supervised by Prof. Mira McWilliams .
We use the Twitter API to collect real time tweets related to Bitcoin
and apply sentiment analysis on them. We then use this data to
predict increase/decrease in Bitcoin's price. Interestingly, volume
of tweets seems to be a better predictor than sentiment in most time
frame.
