Principal Investigator of the INSPIRE project funded by the University of Milan (2018-2019)
The proposed research will focus on automatic systems specifically designed for intelligent signal processing addressing the generic problem of learning evolving probability distributions in non-stationary environments. There exist many applications that generate data from nonstationary environments, where the underlying phenomena change over time. Examples of these applications include making inferences or predictions based on acoustic data acquired for the purposes of speech emotion recognition, analysis of musical signals, bioacoustics (e.g. farm monitoring including assessment of the health of the animals), biomedical applications, including respiratory sound analysis, etc.
Learning in nonstationary environments requires adaptive or evolving approaches that can monitor and track the underlying changes, and adapt the model to appropriately accommodate those changes. After detecting changes in stationarity, such approaches should a) assess the validity of the learnt models, b) gather incoming novel data, c) select from the pool of available and incoming data those which represent best the current conditions/phenomena, and d) finally decide whether to update or discard and replace the learnt models. To this end, we propose to explore a rather promising scientific area called transfer learning which offers the ability, among others, to transfer knowledge between data of existing and novel classes. There, knowledge transfer, if done successfully, would greatly improve learning efficiency since one is able to combine knowledge received online with the available one for constructing a system able to explain current observations.