This project is a continuation of the European MASH project, which will end this December, and aims at pursuing the study of the collaborative development of large feature sets for machine learning.Our motivation is dual.
First, we want to investigate research topics we met during MASH, which we did not anticipate when writing the original project. In goal-planning, we have developed an new mimicking algorithm for POMDPs, which combines dynamic programming and Boosting to automatically decompose a policy into macro-actions. Additional research will allow to incorporate mechanisms from state-of-the-art mimicking techniques for MDPs. In machine learning, we have developed variations of Boosting to deal with very large feature sets. However, we have did not have the resource to investigate on-line training with large training sets. We propose here two novels algorithms to extend our work in that direction.
Second, we want to maintain the existing MASH web-based collaborative platform. It includes tools for the on-line development and test of image features, and teaching-oriented instruments to facilitate its use for lectures in machine learning, programming, or signal processing. We are now in position to advertise the platform to a large public and reap the benefits of three years of development.
The success of MASH, as evidenced by the quality of the developed software platform, and by the resulting publications, demonstrates the validity of its core concept. This proposal is a natural continuation, to build upon the foundations we have laid, and to assess the value of the collaborative design of large image feature sets.