Cardinal: A metrics based Active Learning framework
Data labeling is a tedious yet necessary task to train Machine Learning models. This iterative process alternates data labelling and model training. That’s what Active Learning is tackling. We present “cardinal”, a package that relies on interpretable metrics to help selecting the most representative samples while minimizing selection bias. We prove performances similar to state-of-the-art methods
Machine learning researcher with a taste for unsupervised methods, weak supervision, and human in the loop learning. I did my PhD at Inria on an extension of dictionary learning to multi-subject estimation in large cohorts in order to generate functional segmentations of the brain. I have spent 3 years at Criteo maintaining and developing the recommendation platform. I am now research scientist at Dataiku where I lead the active learning efforts.