Hello,

I'm new here, and I'd like to learn.

TLDR: I would love your opinion and  to discuss about the possibility  of
implementing an intelligent tutor bot. What are the problems? What are the
obstacles? Is the state of the art enough? What subjects need to be tackled
before achieving this?
Can you please help me on this?

Longer version:

I'm working with a group of people to create an intelligent TutorBot that
helps people during learning and guides them thought different materials,
examples and .

I've been looking for technologies that could allow to do this, and last
Thursday I've found a presentation from Jeff Hawkins about CLA and Sparse
coding, and also found about cortical.io After what I was amazed and
completely exited about the technologies.

I've spent all my non-working time last days reading, studying and watching
every single thing that I could find about it. Trying to understand the
concepts behind and trying to figure out if the goal I'm targeting is
achievable in a reasonable time (at least a first technology demo).

After reading the papers at:
http://arxiv.org/abs/1411.4702
http://numenta.com/learn/properties-of-sparse-distributed-representations.html
http://www.eecs.harvard.edu/~michaelm/NEWWORK/postscripts/BloomFilterSurvey.pdf

I've focused mainly on the NLP videos and resources as:

https://www.youtube.com/watch?v=X4XjYXFRIAQ&start=7084
https://www.youtube.com/watch?v=hjMjhhmYKhI
https://github.com/subutai/nupic_nlp
and also the examples and tutorials from the hackathons. I haven't finished
with all the material yet.

I've also started with NuPic installation (I have managed to install it
correctly). Note on installing the software, the bug on numpy version
(conflict 7 and 9) is painful. I wrote a script that solves the issue for
me (I should clean it up and upload it to github).


To go into more depth if I imagine a TutorBot I can see different macro
tasks needed:

   - Guidance to student, for example asking targeted (and necessary)
   questions to the student
   - Remembering what was discussed about, and using those facts in the
   future (continuous learning)
   - training the TutorBot itself


One of the problems is context. I thought about the possibility of
clustering to be able to retrieve subjects that are similar, using
cortical.io API will be even possible to create a signature for phrases and
sentences that are meaningful for the topic in discussion. Later a CLA
trained algorithm can predict the best option for discussion.

There is a difficulty here, that if it learns continuously, how to evaluate
the success of the task?

There are technologies like AIML (it is not intelligent and it is too much
work to be able to create and maintain  a rule based bot) we want to build
something at which I can give knowledge in huge chunks and that it will be
smart enough to use that knowledge when it has to.

One of the thoughts is training the tutorbot with many discussions, and use
the prediction from the CLA directly, for what an action will be required
(motor response maybe?)

About the possibility to train the system, about CLAs I haven't found (or
understood) a way to direct the training of the algorithm with feedback
(positive, and negative feedback), Is this possible to achieve with current
state of the art? If not, what is needed?


Well, I thing is a nice subject to discuss, besides the practical
application I'm looking for, because it implies many other areas where
learning can be used.

Best


-- 
Ing. Leonardo Manuel Rocha

Reply via email to