Fast sequential decision-making via prior knowledge

What is it about?

The goal of automatic sequential decision-making is to learn a machine learning model in order to update knowledge online. However, learning a model from scratch can be time-consuming and inefficient. Consider an online recommendation system that is updated based on the sequence of clients. However, each client has only a few observations. How could we learn a reliable model quickly on a new client? Clearly, we could benefit from previous learned clients' information. In this paper, we present a novel machine learning model for quickly learning new tasks. Furthermore, each task is also learned sequentially, which saves memory and is more efficient.

Featured Image

Photo by Markus Spiske on Unsplash

Photo by Markus Spiske on Unsplash

Why is it important?

Efficiently learning a new task with few observations is critical in many real-world scenarios. In this paper, we proposed various theoretical results as well as a novel practical framework for learning streaming data quickly and reliably. Furthermore, our model could track a non-stationary environment, which is both challenging and promising.