[infurl]

https://www.technologyreview.com/2020/05/06/1001271/an-ai-algorithm-inspired-by-how-kids-learn-is-harder-to-confuse/

In tests with several popular image-classification data sets, the approach almost always led to a final machine-learning model that outperformed one trained by the conventional method. In the best-case scenario, it increased classification accuracy up to 7%.

This makes excellent sense to me. I've always felt that the best way to develop the natural language software that I'm working on would be to start with simpler vocabularies and grammars and layer on more and more advanced features, the same way that kids learn to read in school.

[frankinstien]

That does make sense. What I’ve been working on is a means that works in that identical concept, however, the approach doesn’t use an ANN. The way the machine acquires or I should say improves (its initial vocabulary can be initialized programmatically), its vocabulary is through interaction with humans! The approach can build an information structure that binds concepts with data(experiences).  So there is no need to train on thousands of documents it learns from interacting with a mentor(s). 

[infurl]

What I’ve been working on is a means that works in that identical concept, however, the approach doesn’t use an ANN. The way the machine acquires or I should say improves (its initial vocabulary can be initialized programmatically), its vocabulary is through interaction with humans! The approach can build an information structure that binds concepts with data(experiences).  So there is no need to train on thousands of documents it learns from interacting with a mentor(s).

I'm very glad to hear from another developer who is taking the symbolic AI approach (that is, if I understand you correctly). Machine learning is getting all the attention these days but GOFAI (good old fashioned AI) will still be an essential component of any future AI. After all, our own success is due to our ability to process symbols isn't it. 

[Dee]

AI learns the way kids do, that's very realistic.


However, for AGIs, should AGIs learn (self-learning or being trained) in like 6 years similar to kids
those learn basic things before going to school.
Imagine a train loop that runs for 6 years straight  , seems impractical.


Possible AGIs should self-train, and stay alongside humans to get guidance, this is rather like kids' way.

[infurl]

However, for AGIs, should AGIs learn (self-learning or being trained) in like 6 years similar to kids
those learn basic things before going to school.

It's not the length of time that matters, but the way in which information is introduced. You start with simple words, phrases and concepts, and over time, make them more complex and difficult to grasp.

Think of the difference between the books that you learned to read with, and the books that you read now, and the intermediate levels along the way.

[Art]

I absolutely agree with you infurl. Over many years I've experimented with lots of bots that started out only knowing the rules of the game, so-to-speak. That is, there were no pre-scripted responses for the bot to use. At first, their skill at constructing logical responses was horrible and without reason. Gradually, over time, the bot would learn not only which words went where in the sentence but which ones to use and in what order to use them to best make sense for the user.

Some even longer trials would have the bot communicating in a very human-like fashion. Of course, the conversations weren't completely without error as this was over 2 decades ago and the technology then was certainly not up to speed to where it is now.

Yes, the more data a bot is given or allowed access to and the better the decision algorithms are regarding how and which words or phrases to use, the more positive and less error-prone the bot will become.

We are so close to being there or dare I say, they are so close to being like us that within 10 years or less, they will be indistinguishable from humans. In fact, some are practically there right now.

The human-like android or humanoid form is going to be another issue.

[Don Patrick]

I pretty much have to start with a kindergarten vocabulary, because I've built my program to look at surrounding words to determine the category of an unknown word. After it has categorised a word, it can add it to its vocabulary and in turn will be able to understand words in further sentences that feature the previously unknown word. You can imagine it's hard to guess what a word signifies if you don't know most of the other words in the sentence either.

The same goes for learning facts, which can help out with disambiguation and such. I have a text file with kindergarten level facts about animals, people, items, buildings, and so on, written in very simple language so that I can be sure the program interprets it correctly and builds up a reliable base of knowledge. Whereas if I throw it in the deep end from the start, it'll misinterpret a lot that it wouldn't if I had started off slow. Now that I've begun working on cause and result relations, I'm starting to think that fairy tale books are actually an educational way of teaching children sequences of events without requiring them to fight dragons for the experience.

[frankinstien]

However, for AGIs, should AGIs learn (self-learning or being trained) in like 6 years similar to kids
those learn basic things before going to school.

That is a good point but your assumption is that the way we teach today is effective! I'll give you an example:

I had a girlfriend who had a 4-year-old, he did attend the Montessori preschool so there was some pre-training in reading, writing and arithmetic. We decide to give the boy educational computer games. Many assume that because of the content of the game children are learning the subject matter. What happened was far from learning the subject matter! The boy just wanted to pass the levels of the game and win, that was it, learning wasn't the boy's objective. So the boy would try and solicit our help to win the game. Now most parents would simply give in and say the boy is too young for the subject matter presented in the game, after all, he's just four. There is a book that I had read about ANNs learning grammar called "Rethinking Inateness" by Jeff Elman, etc. I was very surprise to learn that the expectation of ANNs is they are likely to fail and one has to either re-think the data content or train the ANN on a simpler task for it to succeed at the objective. Well, I tried that very approach with the 4-year-old. When he would scream and cry because he couldn't figure out how to pass a level I would look at the material and then restructure an explanation and examples that the boy could relate to. I then tested the boy on a data set that wasn't in the game, just as you would do with an ANN. This actually worked, So I then found some children workbooks for math and science, they were around, a the time, $3.00 apiece, I then set up a monetary reward system for completing the books successfully. If he had problems I would go through the same routine and explain the material in a way the boy could relate to. Within about 6 months, the boy is now five, he's doing addition, subtraction, multiplication, and division on his own! It later improved to long division and multiple digit multiplication which is 3rd to 4th grade math and he's only five.

It wasn't that the boy was gifted in any way it was the material that oftentimes proved difficult because of how it was presented. By figuring out how to communicate the ideas and concepts to the boy so he could relate to the material proved pivotal in his mastery of the subject. So how we teach A,I. is critical in getting A.I.s to master tasks or subjects. So while it takes years to teach humans we can still half that time, at least!  If we think about how to do this with something that can process at the speed of light it shouldn't take anywhere near the time it takes to teach humans, we just really need to think about how to present the material to it...

[yotamarker]

I haven't a clue as to how kids learn, please examplify

[HS]

Here's how I see it... The information has to be layered so that each new layer is directly related to the previous layer. That way you get a neat and complete connectome. Start with the most basic true principles of the world, which agree with the kid's established perspective, and successively add to it without skipping steps. It always relates back to the fun-da-mentals, which is fun, for the mind. Explain new things in terms of old things. Everybody feels the environment, we all have an understanding of direct experience. Hot, cold, soft, sharp, tasty, gross, light, dark, loud, quiet, etc, are direct experience and are the building blocks for modeling objects and concepts a step from direct experience. Then those new objects and concepts can be combined to model the world two steps from direct experience. That’s when the epic music starts playing in the classrooms.

[Art]

I haven't a clue as to how kids learn, please examplify

Come on...you're not that old. Surely you can remember your 1st - 4th grades in school?
How did the teachers "teach"? What material did they give you? Did you do a lot of repetition? Were certain themes often reinforced through examples or experiments? Were you required to read some material out loud (which helps to reinforce for the reader and the listeners)?

Were you (or others) required to memorize certain facts or methods for certain subjects? ex: "I before e except after c"...etc.