Activity Recognition from Accelerometers in Smart Homes
When performing classification tasks such as Activity Recognition in the smart home environment, we are presented with many interesting challenges for Machine Learning: noisy data; missing packets; limited training data; and differing training and deployment settings. We first propose a Bayesian approach to Dictionary Learning (DL), which is shown to be robust to missing data and able to automatically learn the noise level in the data. The coefficients from DL are then used in a classification setting.
Here we take the Bayes Point Machine, a Bayesian classifier, and extend it for use in a Transfer Learning setting, and show how a combination of Active Learning and Transfer Learning are able to efficiently adapt to the deployment context with only a handful of labelled examples.