Towards Instructable Connectionist Systems

At least three disparate channels have been used to install new knowledge into artificial intelligence systems. The first of these is the programmer channel, through which the knowledge in the system is simply edited to include the desired new knowledge. While this method is often effective, it may not be as efficient as learning directly from environmental interaction. The second channel may be called the linguistic channel, through which knowledge is added by explicitly telling the system facts or commands encoded as strings of quasi-linguistic instructions in some appropriate form. Finally, there is, for want of a better phrase, the learning channel, through which the system learns new knowledge in an inductive way via environmental observations and simple feedback information. These latter two channels are the ones upon which we wish to focus, as they are the hallmarks of instructable systems. Most instructable systems depend upon, or at least heavily favor, one of these two channels for the bulk of their knowledge acquisition. Specifically, symbolic artificial intelligence systems have generally depended upon the explicit use of sentential logical expressions, rules, or productions for the transmission of new knowledge to the system. In contrast, many connectionist network models have relied solely on inductive generalization mechanisms for knowledge creation. There is no apparent reason to believe that this rough dichotomy of technique is necessary, however. Systems which are capable of receiving detailed instruction and also generalizing from experience the both possible and potentially very useful.