ABSTRACT
To transition from artificial (narrow) intelligence to artificial general intelligence will require incorporating additional fundamental learning principles that evolved in biologically intelligent systems. One such property is the ability to lifelong learn, that is, to use incoming data to improve performance on essentially all tasks, both past and present, without catastrophically forgetting anything important. We provide a general framework in which an intelligent agent can perform lifelong learning, and then propose a concrete algorithm, generalizing decision forests, to achieve it. Theory, simulations, and real data applications demonstrate the power of this approach.