Why are neural networks always trained "by themselves"?

Question

In the current rush of artificial intelligence research, fueled by NN, independent of the paper I choose, the NN are always trained by themselves. Sure, there are architectures that combine CNN and RNN or LSTMs in a way that can help to solve multiple problems interacting (like labeling images with human readable text snippets), but independent of this, they always learn by themselves.

• Supervised networks just get a bunch of examples to learn from.
• Reinforcement algorithms run around virtual spaces falling over hundreds of times before learning to walk properly.

This might sound silly but no one helps them and no one plays with them. Children, researchers, puppies, dolphins, ... all intelligent beings we are aware of interact with each other. There are many forms of learning, one very important one being social learning which includes observational learning. We imitate, copy and learn from others all the time. There is also a new idea of looking at knowledge that suggests that we all aren't as independent as our culture might have taught us to think we are. That knowledge really lies in the connections between individuals instead of inside of each agent itself, just like the connections of neurons are what makes the brain work, not the neurons themselves.

Couldn't such interaction between learning agents be enormously valuable? If you throw 5 agents in a room (NN a Knowledge Based Agent a simple pre-programmed one etc) and let them learn from each other through observation and inspiration, what effects could one expect?

I am not talking about ensemble learning as this just throws several hypotheses together and takes the average over all results. I am talking about a complex way of agents interacting with each other during learning.

Is there any research on this idea? If so, what were the results? This is the only one I found:

• Learning to Communicate with Deep Multi-Agent Reinforcement Learning

Answer 1

The way children learn is in many ways supervised. It is true that certain abilities are there genetically (visual system, object recognition, to large extent voice recognition), but a lot of human experience is gained as a result of response, either from the environment or from the mentor (parent, teacher, etc).

Social interaction is possible only when a human already learned a lot (things we consider trivial, but in fact pretty complex), in this sense this is equivalent to AlphaGo learning to play Go by playing with a copy network. NNs that can only perform object or voice recognition can't interact, they are too primitive for this. And a nearly all of todays AI is on this stage right now.

However, it is true that NN learning is much less efficient than human learning and current models are far from AGI. I think RL is the closest area that studies real multi-agent environments.