Edwin van der Helm

As a proof of concept for a product, I prepared a large dataset using only publicly available data and trained a neural network for named-entity recognition in medical publications. Because there was no dedicated training data, I had to very carefully combine and augment the data that was available and adjust the neural network architecture to use it effectively.

In a few months, I was able to get results that could compete with models trained on proprietary data that had cost other teams years to compile.

I have designed and implemented an easy-to-use distributed automated hyperparameter optimization library that combines the strengths of different optimization techniques, including HyperBand, Tree of Parzen, regression trees, and random forests.

I led a large distributed team to find and prepare training data on financial news articles. For those articles, I also designed and trained a neural network to find relevant articles based on semantic text content.

I designed and trained a neural network for classifying database tables into several thousand custom classes. Since many of these tables were very large, the data pipeline had to efficiently process numerous terabytes of training data.

Since TensorFlow was first made publicly available, I have been the lead architect and developer of a library for streamlining neural network development for enterprise customers; including training run management, visualization, data processing, augmentation, and hyperparameter optimization.