4
$\begingroup$

A neural network is usually programmed to learn from datasets to solve a specific problem. Essentially, they perform non-linear regression.

Could a neural network be programmed to receive input from a human, like a terminal, to begin to grow and learn (similar to how a child learns)?

A program that neither knows its purpose nor specific data sets but is given enough information to learn based on the input, ponder the input, and ask questions. A child discovers their purpose (in destiny based philosophy) through experience. Thus, could an AI be created that would learn its purpose over time?

Grow both by continued development, maybe adding extensions that add image recognition, speech analysis, etc., and through user interaction. Eventually learning "moral imperatives" or simple the do's and don't's and how to interact with data.

A case scenario would be a Question & Answer session with the neural network and a large data set. Where the human operator knows the answers. At first, the question and the answer are supplied to the neural network. Giving it the ability to find the answer supplied through deep learning. A guaranteed confidence score of (1) - as the question is pondered the closer it gets to the answer the more it "learns".

The next step is supplying the question and waiting for the answer. The human still knows these answers but is testing the "learning machine" to see if it is truly learning and not "repeating the answer". The answer is supplied by the machine and the human returns with either a percentage that the machine is right (hopefully and eventually matching its confidence score). and after an amount of failure provides the right answer to the machine to repeat the first step and improve learning.

The last step is being able to have the machine answer the question with the human not knowing the solution, thus completing the learning cycle. The human would test the solution and report the results to the machine and the machine would adapt the process and continue learning. However, this time it would begin learning from a data set of results. Hopefully learning "data mining" during its question and answer session.

$\endgroup$

4 Answers 4

3
$\begingroup$

In principle, yes, what you are proposing can be done. The exact details of how to do it are an open research question. The details would also depend on exactly what your goals for the system are. If you're just trying to build some domain specific system that learns a very specific kind of knowledge, then that's probably going to be easier than building an AGI that learns like a child does.

What I will suggest, although this is not proven, is that building a really powerful system of this sort will probably require more than deep learning. I would also caution anybody interested in AI against thinking that deep learning is the "be all, end all" of AI techniques. My guess is that doing this well will ultimately require a multi-agent system, maybe something like Minsky's "Society of Mind" approach, or a Blackboard model, with collaborating agents, each specialized for various aspects of intelligence. My feeling is that you will, indeed, need deep learning for classification / pattern matching, but possibly also things like Case Based Reasoning, K-Lines, a Semantic Network, Rule Learning, BDI, and other techniques, working together.

$\endgroup$
2
  • $\begingroup$ Kind of how our "human brain" works with portions of our brain containing self-contained personalities that assist in the task at hand - the evidence is seen by people suffering from a traumatic event that causes multiple personality disorder - a part of the brain resurfaces to solve or protect the primary/real personality. So an AI of smaller AI's that learns and grow on a specific purpose oriented task. $\endgroup$ Jul 8, 2018 at 21:37
  • 1
    $\begingroup$ Example: registering to handwrite, identifying objects, passing data, connecting to the internet, interacting with a human... etc.... I imagine many of them would be very basic layers paired with complicated neural network layers for full computation. $\endgroup$ Jul 8, 2018 at 21:38
3
$\begingroup$

You might also be looking for active learning, where the machine learning algorithm interactively queries the user to label certain unlabeled training examples. Active learning is similar to semi-supervised learning, in which there are labeled and unlabeled examples.

$\endgroup$
2
$\begingroup$

Technological advancement has historically been measured by the processing speed of computing machines. Cogitative behavior psychology is proving to correct human processor disorders such as ADHD, anxiety disorders, addiction, and other psychological disorders preventing humans from normal human interactions and social learning. Cognitive processing speed is very different from the speed per second of programmed computations. Just as humans are learning that a new program for energetic children works better than amphetamines, we will learn a new way to teach AI. AI however is capable of the exponentially growing fast speed of computations per second. AI will indeed and is indeed guiding humanity to base the direction of machine learning along the human path to enlightenment. From providing a computer opponent for chess to face recognition robotics, human learning is based on our amazingly complex brain power and we will always be necessary for advancement of technology.

Parkaire Consultants, (2012, February 24). Cognitive Processing Speed. Retrieved July 7, 2018, from http://parkaireconsultants.com/cognitive-processing-speed/

Staughton, J. (2018, June 21). The Human Brain vs. Supercomputers... Which One Wins? » Science ABC. Retrieved July 7, 2018, from https://www.scienceabc.com/humans/the-human-brain-vs-supercomputers-which-one-wins.html

$\endgroup$
1
  • $\begingroup$ Not always. A well equipped AI, using aspect-oriented programming, or maybe just the ability to recode itself, could learn and advance under human intervention until the AI becomes so complex that only itself can understand it's development. The development of AI is only limited to our current understanding of computing and physics. $\endgroup$ Jul 8, 2018 at 21:40
2
$\begingroup$

It might be possible, but, in my opinion, it won't be very successful, given that you need to somehow specify a purpose, even if that purpose is something like trying to be like humans. The most important thing of an intelligent machine is that it follows a goal, that is the very essence of intelligence.

$\endgroup$
0

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .