Representation learning, the process of extracting relevant features to represent the data, is one of the first steps of machine learning algorithms. This process has been refined over the years and representations are now often learned by powerful deep learning models, stored and reused later on by simpler models for various downstream tasks. However, these representations are not easy to interpret and it is often difficult to understand which latent factors are important for downstream models.
Disentangled representation learning aims to provide representations where each factor is encoded separately, resulting in transparent and simpler representations. The current state of the art disentangled representations are produced by a family of variational auto-encoders. These deep learning models are composed of an encoder and a decoder. The former learns to encode disentangled latent factors and the latter to generate data from this latent representation.
However, disentangled representation learning with variational auto-encoders is still far from solved. Indeed, the models are capable of learning disentangled representations, but their usefulness on downstream tasks is yet to be proven. Moreover, when the architecture of the auto-encoders is powerful enough, the decoder can generate new data without using the latent representation provided by the encoder. Hence, the latent representations are not learned during the training and result in unusable random embeddings.
The goal of my PhD is to investigate how these latent factors are encoded, assess their impact on predictive models, and improve their quality.
I am a member of the following research groups:
My main research interests are disentangled representations learning and variational auto-encoders. I am also interested in information theory, variational inference, and free energy.
I have taught: