How would a human-like artificial visual perception look like? We may not know the answer to this question, but we can forsee such a perception shall have the following characteristics at minimum: it will not demand many examples to learn a new task; it can establish an abstract relationship between learned tasks and transfer knowledge across them; it can accumulate, preserve, and enhance learned tasks. The field of computer vision has undergone a major shift with the dominance of neural networks with a remarkable success at various problems. The vast majority of employed methods use fully supervised learning which essentially aims for a task-specific system utilizing a massive amount of curated training data. Despite their notable success, this approach has prohibitive scalability issues with both amount of required training data and, more importantly, addressed tasks as they fail to efficiently generalize to novel problems.

In other words, they constitute a narrow perception, and not broad. They also bear an unsatisfying sentiment as more efficient alternative approaches to learning are deemed feasible; for instance, cognitive studies propose living organisms can perform a wide range of tasks for which they did not received direct supervision by learning proxy tasks [Held1983,Smith2005,Rader1980]. This suggests achieving success by going outside the paradigm of fully supervising a task-specific system is possible. Therefore, it is natural to invest on alternatives to supervised learning. This is particularly important for the academic community as the experimental trends show common neural networks are nearing the practically satisfactory performance on predefined tasks if supervised with enough training data, and industries are mastering this approach for their needs.