Current machine learning and deep learning models are trained by tasks that come mostly from real-life applications with practical purposes. Few tasks are designed specially and systematically for the purpose of studying pure intelligence. Why is this very important? Much efforts today in the AI community have been put on designing AI models that are AI’s brains, whereas less studies are concentrated on how to design and organize a series of tasks that can help AI grow.
On the contrary, for human intelligence, we do not put too much thoughts on how our brains are built, whereas we would emphasize the importance of education, create educational environemnt and develop curricula as good as possible to make sure our children’s intelligence would grow from a blank canvas into a stunning piece. In this regard, studies on education should be treated as part of the whole ecosystem of AI research. The stories of feral childrens raised by wolves tell us a undeniable fact - the high-level cognitive ability is a production of nurture in education. In other words, education is not only a channel to pass on knowledge from one generation to the next, but also a powerful tool to significantly increase individual cognitive ability.
The term “task system” for machines is derived by analogy to education system for humans. However, we do not refer to the exact meaning and content of human education when mentioning task systems for machines, just as we could not teach a dog the way we teach a human. Instead, machine education should be tailored to suit the need of developing machine intelligence.
A task system should produce a set of tasks, similar in some aspects but different in others. For example, those tasks are set up to solve the same prolem but with different difficulities as in curriculum learning; those tasks face the same prediction goal but with different input domains as in transfer learning. However, lacking a top-level world design, these sort of task setups are not comparable against the real-world ones for humans and even for animals. Therefore, a well-designed task system should be built on a world, a common environment base, which all tasks should use the materials of as their design elements and also be consistent with. This world does not have to be exactly identical to our real world. However, the world should be able to grow with new elements and details added, so that new tasks could be continuously generated from it.
For the learning purpose, there is a more profound goal for machines than simply learning to solve one specific task, which is to estabilish a deep and full understanding of how the world works through attempting to solve tasks one after another. I list several types of tasks as follows:
- Predicting unseen states or things in the world based on partial observation. Machines need to develop this kind of ability to break the barrier of time and space, and also overcome the limitatin of its own body such as sensors.
- Causal reasoning about the inner workings of the world such as the physical laws. Machines should be aware of the possible consequences of each action as well as an occuring event by watching real scenarios, imagining counterfactual scenarios and intervening in the world. This ability will help machines develop its own mental environment model to mimick and interpret the world it lives in.
- Planning. Based on the environment model developed by causal reasoning, machines should be able to do planning and navigate itself in the action space to finish a relatively complex job.
- Communicating and collaborating with other machines. Intelligence emergence is a production of a society rather than a single individual.
- Developing a set of abstract symbols with some syntax within a society to boost the intelligence level of the entire society. This is the ability of acquiring languages, including assigning a symbol a specific meaning, using more symbols syntactically to capture or constuct more complex meaning, understanding symbols generated by other machines, and knowing how to behave under the guidence of symbols.
- Continously developing abstract symbols into a more adavanced language system. Ultimately, machines own the ability to write novels, summerize life experiences, and develop their own strategy and philosophy individually and globally.