Task-Nonspecific and Modality-Nonspecific AI.

It is widely accepted in Artificial Intelligence (AI) that different tasks require different learning methods. The same is true for different sensory modalities. However, auto-programming for general purposes seems to require a learning engine that is task-independent and modality-independent. We provided the Developmental Network (DN) as such an engine to all contestants of the AI Machine Learning Contest 2016 for learning three well-recognized bottleneck problems in AI—vision, audition, and natural languages. For vision, the network learned abstract visual concepts and their hierarchy with invariant properties and autonomous attention. For audition, sparse and dense actions jointly serve as auditory contexts. For natural languages, the network acquires two natural languages, English and French, conjunctively in a bilingual environment (i.e., patterns of text as inputs). All the three sensory modalities used the same DN learning engine, but each had a different body (sensors and effectors). The contestants independently verified the DN’s base performance, and competed to add (hinted) autonomous attention for better performance. This seems to be the first task-independent and modality-independent learning engine, which was also verified by independent contestants. Much remains to be done in the learner-age related sophistication of learned tasks.