In order to increase flexibility of the VAE we proposed to utilize the idea of the normalizing flow. First, we used a series of Householder transformations to obtain a deep structure of latent variables. Next, we extended the linear Inverse Autoregressive flow by using a convex combination of lower-triangular matrices. Further, we proposed a non-linear normalizing flow by utilizing the Sylvester's theorem. The main idea of the approach was to parameterize weights of the transformations using orthogonal matrices by applying Householder matrices, a numerical procedure and a permutation matrices.
In the VAE framework the Gaussian distribution is a default option for both the prior and posterior. However, we hypothesized that this could fail for different latent topologies, especially for a latent hyperspherical structure. To address this issue we proposed to use a von Mises-Fisher distribution instead. Through a series of experiments we showed how such a hyperspherical VAE is more suitable in discovering a latent structure and it is able to outperform a vanilla VAE on the image dataset and citation network datasets.
Our next approach was to extend the VAE framework by using a new type of prior ("Variational Mixture of Posteriors" prior, VampPrior). The VampPrior consists of a mixture distribution with components given by variational posteriors conditioned on learnable pseudo-inputs. We further extended this prior to a two layer hierarchical model and showed that this architecture learns significantly better models. We provided empirical studies on six image datasets and showed that our approach delivers either the best results or performs on par with state-of-the-art results on all datasets. Next, we utilized the VampPrior in the fair classification setting. Fairness is a statistical property that is very important in many practical applications including medicine. We proposed a two-level hierarchical VAE with a class label and a sensitive variable.
Training a whole slide imaging tool requires relatively large amount of computational resources and providing pixel level annotations is extremely time-consuming. In order to overcome these issues, we proposed to apply multi-instance learning combined with deep learning to histopathology classification. Our goal was to utilize weakly-labaled data to train deep learning models in an end-to-end fashion. We discussed different permutation-invariant operators and proposed a new one basing on the attention mechanism. We applied the newly developed techniques to four cancer data, namely, breast cancer, colon cancer, esophagus cancer and prostate cancer. The obtained results were of great clinical potential.
Beside the international conferences and two seminars at the University of Amsterdam, the project was dissemination through different channels and media. I had invited talks at the Summer School on Data Science (Split, Croatia, 2017) and in multiple institutions (CERN, CWI in Amsterdam, TU/e in Eindhoven, MPI in Tuebingen). Moreover, I took part in Open Day at Science Park (Amsterdam, 2017), one of the biggest events for participants in every age. I gave also three interviews (on Polish radio, on Polish TV, in Polish “Pryzmat” magazine) and one short interview for a Dutch university magazine. Additionally, the DeeBMED was described in the newsletter of the Wroclaw Center of Technology Transfer. Last but not least, I have launched a Twitter account where I included my thoughts and successes of the project (observed by >1000 users), a GitHub account (followed by >80 users) and a project’s website (
https://jmtomczak.github.io/deebmed(öffnet in neuem Fenster)). The dataset developed and used during the project is publicly available (
https://zenodo.org/record/1205024#.W6_oBnUzbCI(öffnet in neuem Fenster)) and was viewed ~280 times and downloaded >30 times.
