Life-long Learning Through Task Rehearsal and Selective Knowledge Transfer
Daniel Silver, Robert E. Mercer · BiblioBoard Library Catalog (Open Research Library) · 2009
The following summarizes the key observations that have been made while developing the TRM with MTL prototype system and conducting the experiments presented in this paper. A number of avenues for future work are suggested. 6.1 Retention and generation of task knowledge. The experimental results demonstrate that TRM has the ability to retain accurate task knowledge in the form of neural network representations. The experiments also show that the TRM can generate virtual examples with the same level of accuracy as the retained network hypotheses. Selective Retention of Accurate Task Knowledge. The TRM prototype system has demonstrated the ability to selectively retain only hypotheses which have met a pre-defined level of generalization accuracy based on the hypotheses classifying independent test sets of examples. This ensures that a level of domain knowledge accuracy is upheld. Efficiency and Scalability of Task Knowledge Retention. The experiments demonstrated that TRM provides an efficient storage of task knowledge. The hypothesis representations saved in domain knowledge implicitly retain the information from the training examples in a compressed form. In the case of the Logic domain the 8 tasks have a total of 325 actual