Online Budgeted Learning
The emergence of large scale databases and big data has recently motivated the development of budgeted machine learning models able to learn under operationnal constraints in term of memory/CPU consumptions, data access, etc… It involves integrating the constraints of scarce resources in the learning process itself. In parallel, in the neuroscience community, reinforcement learning online capabilities are understood as results of the coexistence of complementary learning systems, also based on limited budgets (mainly in terms of computational cost). Based on the observation that the recent context of learning under stress is highly relevant both for massive data processing and neuroscience, we aim to study this issue in the online learning context that seems suited for these two families of problems. We seek cross-fertilization between the two fields: 1) import the multi-model from neuroscience in statistical machine learning architectures as a potential solution to budgeted data analysis and prediction; 2) update the reinforcement learning concepts in use in neuroscience by a confrontation with modern budgeted learning frameworks.
Please follow the link below to have more informations: http://onbul.lip6.fr/