Learning can happen through various means like supervised, unsupervised, reinforcement, incremental, inductive, domain, knowledge based etc. Building a new machine learning system today is more than just the learning. It should have learning capabilities persisting for lifelong.
A Lifelong Machine learning system, as defined already is a system that learns many tasks over lifetime from one or more domains. They effectively and efficiently retain the knowledge they have learned and use that knowledge to more efficiently learn new tasks. The essential ingredients of this system are the retention of the learned knowledge, the selective transfer of prior knowledge when learning new tasks and a systems approach that ensures the effective and efficient interaction of the retention and transfer.
When we opt to build such system, following are the challenges:
- Best representational space for new knowledge
- Tracking universal and domain knowledge
- Integrating new knowledge
- The acquisition, representation and transfer of domain knowledge
- Meta data for cross domain communication
- Transferring the knowledge to the new task through the rehearsal of previously learned tasks
- Representing the knowledge in the best available functional form (eg. as training examples)
- Representation of hypothesis
- Searching for the best hypothesis within the current representational space
- Building hypothesis, strengthening it and selecting the most related prior knowledge
- Developing a new updated hypothesis
- Expert system to improve learning accuracy
- Continual tracking of a solution so that to achieve a better performance than learning a solution from prior knowledge
- Building agents so that they learn, die and give many new born
- Building an Inductive bias (constraint on a learning systems hypothesis space, beyond the criterion of consistency with the training examples)
Inspired and Referred from Silver, Q. Yang, and L. Li. AAAI Spring Symposium: Lifelong Machine Learning, volume SS-13-05 of AAAI Technical Report, AAAI, 2013.