Adaptive data collection for accelerating discovery rates.
The standard machine learning methods, called "supervised learning", take in a dataset passively and then build a model that can make accurate predictions for future data. In many situations, however, we can choose actively which data to collect (or desire to do so to maximally use the budget). That is, we may collect data wisely (e.g., adaptive experiments) so we use significantly less data while achieving the same performance (e.g., identification of interesting genes). At the same time, adaptive data collection means that we are breaking the standard i.i.d. assumption on the data, which is a significant challenge as theorems and principles developed for supervised learning do not apply here. In this talk, I will talk about novel adaptive data collection and learning algorithms arising from the so-called multi-armed bandit framework and show their theoretical guarantees and their effectiveness in real-world applications. Specifically, I will first show how biological experiments can be performed with a significantly less budget by adaptively selecting what experiments to run next. And then, I will talk about how to accelerate drug discovery rate with bilinear models where we use drug and protein information to guide the search.
Throughout the talk, I will show various other practical applications where bandit methods have proved valuable including AlphGo, hyperparameter tuning, and cartoon caption contest.