Robot Not a Dev? Buy Now
Not a Dev?

Misty Community Forum

Learning in Robotics & Approaching Comprehension

This article presents a wonderful set of experiences from researchers who study infants and children, and it discusses parallels in robotics.

A part of the thesis presented therein discusses what is termed “the six Cs of modern learning: collaboration, communication, content, critical thinking, creative innovation and confidence.” To paraphrase, the learning process for human children goes beyond decoding symbols and making connections, as it is a complex process that is continuous, communal, interactive, and social.

At the end of the article, they conclude that robots can never be used to replace learning in humans, and robots might be destined only to augment human learning ability. For a robot to achieve true understanding, it would need to transcend beyond abstraction and anything formulaic to become “embodied, emotive, and subjective,” because human children don’t just comprehend when they learn; they also create new meaning as they learn.

The article discusses some excellent lessons that we can apply to robots based upon how children navigate their complex world during the formative first few years. Some of my favorite lessons from the article include:

  • Behavior encodes rich information that should not be ignored; it is an indicator that the being is trying to tell you something.
  • Creative play, freedom from constraints, and allowing risk-taking in learning creates richer outcomes and advances the learning process further than learning with the absence of these qualities.
  • Don’t underestimate the power of exploration and it’s impact on the learning process.
  • “Shared attention is the starting point of conscious human learning.”
  • The human brain forms itself through interaction with sensory experience.

Based on your own personal experiences with children and watching them grow up & learn in the world, what takeaways or lessons learned would you propose for people who are developing “learning” algorithms for robots? Do you have any thoughts or ideas about how children learn that we should apply to robots?

3 Likes

Fabulous article. From my observations as a parent, what’s very important for children’s learning is what I guess I’d call “spiraling iterative repetition”.

No one understands something fully the first time they encounter it. Yet everyone is bored and tunes out if they are repeatedly exposed to things precisely the same way.

The solution (and what often happens naturally in childhood exploration) is that the same thing is encountered multiple times from different “angles”, ideally ones that spiral upward in complexity. It may be a weird example, but let’s take doors/doorways:

  • A baby is carried through doorways.
  • A baby is carried while their parent turns a doorknob and opens a closed door.
  • A baby hears something on the other side of a closed door.
  • A toddler reaches for a doorknob and discovers it’s hard, smooth, and moves a bit.

Each of those observational encounters iterates differently on doors/doorways. Each of them is incredibly simple, so simple that it is just part of the unconscious construction of a map of how the world works. But, from those encounters, the child has started building their notions of doors, doorways, doorknobs, object permanence (things behind doors are not gone), noise changes due to distance and obstruction, etc.

This is how natural childhood learning works. Now, as a parent, I can also help things become conscious learnings by naming and narrating the child’s object encounters, actions, and reactions, but that’s not needed for the basics to occur.

1 Like

I think it goes beyond multiple experiences from different angles, because that is essentially what we do for face recognition, object detection, etc.

Robots cannot transfer skills and learning to novel problems in the same way that humans can. A human may have never encountered a door with a round knob for a handle, but can figure it out. Robots do not have the “figuring it out” portion.

A good example of this is alpha Go. It is better than the best human at a 19x19 Go board. If you change the board to 18x18 or 20x20, humans could adjust fairly well. Alpha Go would have to be completely retrained.

1 Like

Ah, that does make sense. When a child (or adult) human gets more of those different “angles”, they don’t just form additional instances of truth, they form more powerfully generalized rules. A robot just adds on additional instances.

Eh, I don’t think that’s entirely true @Scott. There’s the whole area of transfer learning, which is specifically aimed at enabling robots to say, learn how to open a door with a round handle when they’ve only ever seen long handles before. Certainly not a solved problem yet, but people are working on it.

And as for ‘figuring it out,’ isn’t that what reinforcement learning is all for? Tell the robot that the door needs to be open, and it’ll just start trying things. It may take forever, but eventually, it’ll figure out some way to make the door be open.

5 Likes

I really like this topic and think this (the quote above) is the right question. The subjects in the article, from my take, seemed to be trying to teach a machine the way we teach a child. But I think the more appropriate idea is to study how a child (or any intelligent being) learns and try to build a model to that.
I am working on a short eBook that hypothesizes: the next big breakthrough in “artificial intelligence” will occur when machines are able to mathematically manipulate concepts. They must be able to (1) create concepts based on perceptions and previously established concepts, (2) retrieve concepts based on context, and (3) apply concepts in order to solve problems (and this application is a mathematical process).
Looking forward to hear what others are thinking.

2 Likes

+1 for the creation and application of concepts/rules being key…