If a wearable device can register what the wearer is currently doing, it can anticipate and adjust its behavior to avoid redundant interaction with the user. However, the relevance and properties of the activities that should be recognized depend on both the application and the user. This requires an adaptive recognition of the activities where the user, instead of the designer, can teach the device what he/she is doing. As a case study, we connected a pair of pants with accelerometers to a laptop to interpret the raw sensor data. Using a combination of machine learning techniques such as Kohonen maps and probabilistic models, we build a system that is able to learn activities while requiring minimal user attention.