It would appear that artificial intelligence will soon be going to school.
Researchers are devising new machine learning techniques that involve what’s termed as “grounded language.” Deep learning today involves feeding an algorithm a lot of data and allowing the algorithm to identify patterns in that data. Some computer scientists are investigating the idea of having algorithms investigate a simulated environment, on the assumption that deep learning occurs when students are permitted to explore an environment. Language and words would no longer be merely abstract entities to the algorithm, but would be associated with and grounded in actual “experiences,” a process that mimics the way humans learn.
Yoshua Bengio, one of these researchers, describes an approach called “curriculum learning:”
The idea is that we don’t just show all the training examples as one big pile in an arbitrary order. Instead, we go through examples in an order that makes sense for the learner. We start with easy things, and once the easy things are mastered, we can use those concepts as the building blocks for learning slightly more complicated things. That’s why we go through school, and why when we are 6 years old, we don’t go straight to university. This kind of learning is becoming more important in training computers as well.
– Yoshua Bengio
As of this moment, the use of the words “school” and “schooling” appear to be metaphorical; no one is talking about a brick-and-mortar school for artificial intelligence. But I wonder if in the future such formal institutions for educating artificial intelligence will emerge alongside schools for children.
GoodAI is a company seeking to build a general artificial intelligence software program. General artificial intelligence refers to a still-theoretical algorithm that possesses the same level of intelligence as a human, in that it can learn any number of tasks that a human can. Current AI can be programmed to learn specific, discreet tasks, like the Alpha Go algorithm that has become world-class (indeed better than any human) at that ancient board game. Watson can be instructed to win at Jeopardy, to analyze millions of medical images, even do our taxes for us. But each task is discreet, directed by humans; Watson does not awake each morning and ask, “What should I do today?” Human intelligence includes our ability to engage any number of intellective activities, hence the term “general.” General artificial intelligence is a kind of Holy Grail for artificial intelligence research, and deep learning is a promising path toward realizing that goal.
In order to achieve GAI, GoodAI is sending its algorithms to school. In addition to programming skills and intelligence into the algorithm, GoodAI “expect the AI to be able to learn. We will teach the AI new skills in a gradual and guided way in the School for AI which we are now developing.” The language GoodAI is employing to describe their plans is very telling. “In School for AI, we first design an optimized set of learning tasks or a ‘curriculum.’”
The curriculum teaches the AI useful skills and abilities, so it doesn’t have to discover them on its own. Next, we subject the AI to training. We use the performance of the AI on the learning tasks of the curriculum to improve both the curriculum and hard-coded AI skills. For teaching the AI, we have created a simulated visual toy world with simplified physical laws. We are designing our curriculum to teach the AI from the most basic rules of the world to the most complex ones, up to the point where it can start learning on its own. The goal is not to teach the AI any arbitrary and specific facts about the world, but the opposite: to teach it useful and general skills for a more efficient understanding and exploration of the world, and for better and more general problem-solving.
I wonder if a new occupation is dawning: the teacher of artificial intelligence. It might be possible that education schools will begin to train teachers who would work exclusively with AI. It is also possible that schools will be redesigned such that artificial intelligence and children will learn together, architected to provide a curriculum of experiences to both artificial intelligence and to young humans.
Even if AI and children are not schooled together, the curriculum-based approach to teaching AI might have effects on human learning. We might view experiential learning as the best way to teach children. If grounded language is the ideal way for AI to learn, might we reimagine schooling for human children to be similarly grounded and experiential? We might see a resurgence in Dewey-ism or Montessori-ism.
When developing the School for AI, GoodAI encountered a problem with broad ramifications: “How should we specify the tasks for the AI?”
When there is no or very little common language, it is very challenging and time-consuming to explain tasks to the AI. For this reason, we are focusing on early language acquisition. To cut down on AI development time, we want to be able to efficiently communicate with the AI as soon as possible.
Early language acquisition is, of course, a critical task in the education of children. Does this mean that tech companies will begin to raid education schools for their talent, rather than seeking such talent in cognitive science, engineering or computer science departments?
More broadly, GoodAI are talking about the need to develop a lingua franca between humans and artificial intelligence, as is needed when two cultures encounter each other for the first time. The next great challenge may well be developing the capacity to translate between the two intelligences.