Projects and coursework from EPFL, spanning NLP, biomedical ML, and graph neural networks.
CS-552 Modern NLP, 2023 | with G. Khurana, K. Naumova
Digital educational assistant built with a full InstructGPT-inspired pipeline: data collection (10k+ MCQA interactions), reward model training (ELECTRA-Large, 90.1% accuracy), supervised fine-tuning (FLAN-T5), and RLHF via PPO.
Key finding: RLHF with small models (80M params) produces degenerate outputs due to distribution mismatch between reward model training data (ChatGPT) and SFT model outputs — a documented negative result on the limits of alignment at small scale.
CS-502 Deep Learning for Biomedicine, 2024 | with P. Lardet, K. Mashimo
Implementation of RelationNet for few-shot learning on biomedical datasets (Tabula Muris gene expression, SwissProt protein sequences). Benchmarked against MAML, ProtoNet, and MatchingNet within a full experimental framework.
CS-503 Visual Intelligence: Machines and Minds, EPFL | with A. Ganuza Izagirre
Study of Vision Transformer robustness against corrupted inputs. Implemented ViT, RVT (Robust Vision Transformer), and a custom FoodViT variant, benchmarking on CIFAR-10/100 and FOOD-101 with their corrupted counterparts (CIFAR-C, FOOD101-C). Explores convolutional patch embeddings, relative/rotary position encodings, and locality inductive biases for robustness.
Neural Interfaces, EPFL | with R. Benichou, C. Flipo, E. Shegurova
Technical report on a subretinal prosthesis design for AMD and retinitis pigmentosa. Reviews state-of-the-art retinal implants (Argus-II, Polyretina, Alpha AMS/IMS, PRIMA) and proposes a device using 3D pillar electrodes to reduce stimulation thresholds and improve visual acuity, addressing effective stimulation, high resolution, and selective cell activation.
DLIB, 2023
From-scratch PyTorch implementations of GraphConv, GraphSAGE, and GAT with multi-head attention. Includes 5 aggregation strategies (Mean, Sum, SqrtDeg, MaxPool, LSTM) and graph-level pooling for classification.
Additional lecture materials, exercises, and homeworks are in coursework/.