Title : Linkflows: a new model for scientific publishing at a granular level
| Presenter | Cristina Bucur |
| Abstract | In this research we want to determine if a fine-grained model of the scientific publishing workflow can help us make the reviewing processes more efficient and more accurate. For developing this model called Linkflows, we are using existing ontologies like the SPAR suite ontologies, the FAIR Reviews ontology and others such as PROV-O and the Web Annotation Data Model. Our contributions include a novel model, Linkflows, that combines these already existing ontologies together with new classes and properties that are able to provide a more detailed, semantically rich view on the reviewing process. We evaluate the efficiency and accuracy of applying the Linkflows model in the reviewing context on a manually curated dataset from several recent open peer review computer science journals and conferences. We perform an initial analysis on the reviews in the dataset considering the Linkflows model and then perform a user experiment where we compare experts with the actual peer-reviewer answers when using Linkflows. The results of this user study are preliminary, and we want to see if we are able to express at a finer-grained level the reviewing workflow and the changes an article or a scientific text snippet undergoes. We also want to see how multiple automatic lexicon-based sentiment analysis tools perform when applied to scientific reviews. Our initial findings suggest that these sentiment detection tools are worse than experts, which are themselves not perfectly aligned with the ground truth. We also notice interesting correlations among different finer-grained aspects of the reviews when using the Linkflows model. |
Title : On equivariance properties in Deep Neural Networks
| Presenter | David Romero |
| Abstract | During the last decades, Deep Learning (DL) architectures have steadily improved state-of-the-art results in a large corpus of tasks in the machine learning field. Although unknown for many practitioners, one important reason for this behavior is the property of equivariance, which has been implicitly (some times even unaware) hard-coded in new upcoming structures (e.g. Convolutional Neural Networks). In this talk we will cover the basics of equivariance and propose a method to detect and overcome undesirable situations that appear when utilizing equivariant networks on non-fully equivariant datasets and to (hopefully) utilize computational resources more efficiently. Furthermore, we will open a brief discussion about the possible inclusion of equivariance in other DL architectures and for other machine learning tasks. |
Title : How informative is the latent space of variational autoencoder when it comes to single cell analysis?
| Presenter | Chao Zhang |
| Abstract | Variational Autoencoder (VAE) is a generative model from the computer vision community; it learns a latent representation of the images and generates new images in an unsupervised way. Recently, Vanilla VAE has been applied to analyse single-cell datasets, in the hope of harnessing the representation power of latent space to evade the “curse of dimensionality” of the original dataset. However, some research points out that Vanilla VAE is suffering from the issue of the less informative latent space, which raises a question concerning the reliability of Vanilla VAE latent space in representing the high-dimensional single-cell datasets. Therefore a study is set up to examine this issue from the perspective of bioinformatics. This paper confirms the issue of Vanilla VAE by comparing it to MMD-VAE, a variant of VAE which has overcome this issue, across a series of mass cytometry and single-cell RNAseq datasets. The result shows MMD-VAE is superior to Vanilla VAE in retaining the information not only in the latent space but also the reconstruction space, which suggests that MMD-VAE be a better option for single-cell data analysis. |