Deep Learning in a Nutshell: Reinforcement Learning

Originally published at:

This post is Part 4 of the Deep Learning in a Nutshell series, in which I’ll dive into reinforcement learning, a type of machine learning in which agents take actions in an environment aimed at maximizing their cumulative reward. Deep Learning in a Nutshell posts offer a high-level overview of essential concepts in deep learning.…

I'd like to apply this kind of deep machine learning to a problem in human education: the matching of learners' characteristics (background knowledge, learning styles and goals) with learning objects (more properly labeled "teaching objects" including Open Educational Resources, free-lance teachers and communities of practice). The machine (AI) would scan the profile of the learner and search online for the most appropriate collection of teaching objects to meet the learner's goals given his/her abilities and background knowledge. Is anyone out there working on this? Please contact me.

Liza, the type of problem you are looking to solve is a bit different to the thrust of this article (Reinforcement Learning), which considers how to learning a task based on the feedback received from repeated attempts/experience (e.g. playing a game over and over). The problem you pose could be addressed with Machine Learning IF there was sufficient data that could be analyzed that links the learning techniques that have worked for people with different learning characteristics. "sufficient" here typically means a LOT. Various techniques can then be used to determine any pattern between the learner characteristics and the most suitable learning resource. If such data exists or can be obtained then I'm sure someone could/would look into it...

Thanks for the comments, Carl. I don't think the data we need exists yet in useable form because a) we don't aggregate user responses to the learning objects they use and b) we don't have well-elaborated learner profiles. But ya' gotta' start somewhere. I'm starting by writing about the idea. Hopefully someone will be able to provide the technical expertise and facilities to begin data collection. I know a lot about cognitive learning characteristics and how measure them. Others will have to throw in the AI and computing power. We'd have to start small and develop a robust matching model. With those two in place there are plenty of learners out in the world to let the machine rip. Because personalizing instruction is such a core issue in education I think this research is fundable -- with the right team. Are you ready to come on board?

It's on the Nvidia site already, anyway I guess the idea is I have is to evolve an aLife using a deep neural network and "soft" associative memory (AM.) The AM only needs an approximately similar input to what it has seen before to produce a meaningful output. If you don't overload the memory it has repetition code error correction. Overall that should make it easier for the deep neural network to learn how to use it.

Then it should be able to learn any algorithms it needs to survive.

I'm working on how to link the two together at the moment.