Calvin Seward wants to know is that the best we can do? We do search, but could we do it better. We do recommendations, but could they be better? It’s a hard question to answer, since we can’t know for sure what the customer would have done if she’d been shown different search / recommendation / pricing results.
Therefore, he’s applying his Mathematics background and strong publication record in image generation to a three prong-strategy: First, he’s collaborating with Ingmar Schuster and Kashif Rasul to improve our probabilistic time series forecasting methods. Then armed with such modeling tools, he is using Thompson sampling to derive policies which harness past data and explore into the future, ensuring policies learn and always get better. Such policies which explore have already been integrated into our outfit generation and recommendation tool. Finally, to ensure such policies work well in the ever adapting and dynamic fashion world, he is actively developing methods for robust reinforcement learning algorithms [link to the new project page], generating policies which are proven to work in shifting environments such as the fast paced, ever changing world that is Zalando.
Current research interests:
Deep Learning, Unsupervised Learning, Object Localisation, High Performance Computing