Deep generative modeling and NLP for finance @ Sony Computer Science Labs (Tokyo, 2022)
I developed a method to compute time-dependent loadings in security factor models based on the optical-flow algorithm and trained a VAE to learn and visualize low-dimensional fund-month exposure embeddings, to support GPIF activities. I also researched NLP algorithms (including approaches using transformer-based language models such as BERT and MPNet) for various problems, such as fake news detection and greenwashing evaluation, to help Nomura Asset management assess corporate sustainability.
Cine-MR image forecasting with RNNs and transformers for latency compensation in radiotherapy @ The University of Tokyo (2017-2021, independent work until 2026)
I designed an algorithm to forecast future frames in dynamic chest and liver MR images, based on motion modeling with PCA and temporal dynamics prediction with RNNs trained online and transformer encoders, for safe MR-guided radiotherapy. The proposed approach is modular, interpretable, and can help overcome data scarcity, motion irregularity, and distributional shift between the training and test datasets for population models.
External marker position forecasting using RNNs trained online for safer radiotherapy @ The University of Tokyo (2020-2021, independent work until 2025)
I proposed implementations of online training algorithms for RNNs (UORO, SnAp-1, and DNI) with closed-form simplifications as well as improved accuracy and reduced time/memory complexity. I assessed their ability to forecast the positions of markers on the chest and abdomen for latency compensation and robotic control in radiotherapy for the first time. In our experiments, modern online learning algorithms enabled RNNs to adapt to irregular breathing patterns with limited data and achieve competitive accuracy compared to prior deep learning approaches trained on larger datasets, while keeping inference time relatively low.
UORO article:
Follow-up article (covering also SnAp-1 and DNI):
Simultaneous segmentation and recognition of characters @ Idemia (Île-de-France, 2016)
I prototyped an OCR pipeline in C for degraded text images, consisting of three steps: over-segmentation, character recognition, and shortest-path graph decoding via the Viterbi algorithm to select the most likely string (cf the image above). It achieved an accuracy 13.4% higher than Tesseract 3.04, an open-source software. The first step of the pipeline (candidate cut generation) relied on the Dijkstra algorithm, as each pixel was assigned to a node in the graph representation of the image and weights were based on pixel intensities; I optimized speed using a binary-search-style approach. Regarding the second step, I implemented from scratch a character-level neural network classifier, trained it on a database of ~20k character images, and improved recognition accuracy by ~5% by including statistical information on aspect ratio.
Graph matching for computer-aided neurological disease diagnosis with PET imaging @ Fresnel Institute (France, Marseille / 2016)
I applied spectral graph theory to the classification of PET images to assess whether a subject's brain is healthy or not. The F1 classification score achieved was 0.78/1.0