Alice Bizeul

I am a doctoral student at ETH Zürich and an ETH AI Center fellow, advised by Prof. Julia Vogt and Prof. Bernhard Schölkopf. In parallel, I am currently a research intern at Apple MLR working on multi-modal generation. In 2024, I interned at Amazon Science in Tübingen, Germany, where I worked on diffusion model fine-tuning. Prior to joining ETH, I graduated from EPFL after a master’s thesis at the MIT under the supervision of Prof. Satrajit Ghosh. At MIT, I worked on the generation of 3D brain scans using GANs.

I am broadly interested in representation learning, (probabilistic) self-supervised (SSL) methods and generative modelling, more specifically understanding how current methods work to then leverage this knowledge to build general representations.

During my PhD, I have been thinking about how we learn representations from observations in unsupervised settings. While discriminative methods (e.g., SimCLR, DINO) generally outperform reconstruction/generative ones (e.g., VAE, MAE), they often rely on less principled approaches. Instead reconstruction-based approach carry their load of challenges, but are driven by principled objectives and require less prior knowledge about the downstream tasks of interests.

Early on during my PhD, I used identifiability and causal representation learning to better understand discriminative SSL in multi-modal settings. Later on, I explored latent variable models as a mean to explain SSL and unify discriminative and generative methods. More recently, I’ve explored Masked Autoencoders and proposed the space of principal components as a more principled and effective masking space for reconstructive SSL. Finally, I am currently looking into the training dynamics of discriminative SSL, specifically how data characteristics affect these dynamics.

More information about me can be found in my resume