Claudia Vanea | Searcharxiv

arXiv2022

A New Graph Node Classification Benchmark: Learning Structure from Histology Cell Graphs

We introduce a new benchmark dataset, Placenta, for node classification in an underexplored domain: predicting microanatomical tissue structures from cell graphs in placenta histology whole slide images. This problem is uniquely challenging for graph learning for a few reasons. Cell graphs are large (>1 million nodes per image), node features are varied (64-dimensions of 11 types of cells), class labels are imbalanced (9 classes ranging from 0.21% of the data to 40.0%), and cellular communities cluster into heterogeneously distributed tissues of widely varying sizes (from 11 nodes to 44,671 nodes for a single structure). Here, we release a dataset consisting of two cell graphs from two placenta histology images totalling 2,395,747 nodes, 799,745 of which have ground truth labels. We present inductive benchmark results for 7 scalable models and show how the unique qualities of cell graphs can help drive the development of novel graph neural network architectures.

cs.LG↗

arXiv2020

Training Ethically Responsible AI Researchers: a Case Study

Ethical oversight of AI research is beset by a number of problems. There are numerous ways to tackle these problems, however, they leave full responsibility for ethical reflection in the hands of review boards and committees. In this paper, we propose an alternative solution: the training of ethically responsible AI researchers. We showcase this solution through a case study of a centre for doctoral training and outline how ethics training is structured in the program. We go on to present two second-year students' reflections on their training which demonstrates some of their newly found capabilities as ethically responsible researchers.

cs.CY↗