Artificial Intelligence used to provide a 10X improvement in personalized education using ‘little data’

So we have an interesting problem that I want to get some advice on:

Background: We are a nonprofit built by a bunch of tech startup founders and enthusiasts that want to do something for our kids: We are building an educational platform, initially for the poorest people on the planet, who don't have access to schools, teachers or the internet, let alone basic infrastructure. A test was done 18 months ago when a box of 40 tablets that were pre-installed with hundreds of educational apps and were given to a remote village in Ethiopia who were 100% illiterate. This box of tablets were left there without any guidance or training on how to use them. The village had a large group of children who had never seen modern technology let alone heard the English language however within hours the first child figured out how to turn one of them on. A few minutes later he taught the rest of the kids how to turn theirs on. 2 weeks later the kids were using on average 57 apps each, a few more weeks after that, they were singing the alphabet song and starting to learn how to read and write letters. After 5 months they learnt how to hack Android.

Recently the Global Literacy Project expanded on this test and deployed an additional 600 tablets to 8 locations around the world and they are getting back some really great results. We are now convinced that this approach can work and we are taking it to the next stage by developing an AI —amongst other things— to guide the child through a completely personalized learning experience that dynamically unfolds as the AI learns more about them. The trick is we need to AI to be on-device as we cannot rely on internet connectivity or the creation of local computing nodes. In addition, because we are ultimately also building it for our kids we don't want their highly personal data sitting on some server or being transferred over the internet where it could be intercepted.

So the problem is this: We want the on-device AI to build an increasingly large and personal data set about the learner (E.G. Learner is X age, has X interests, is good at X,Y,Z, is not so good at A,B,C, Has personality traits Q,R,S, works better with X learning style, has completed T,U,V modules ….). In addition, the content the learner is learning from will have an increasingly detailed list of metadata associated with it. (E.G. Content Type: Video, Topic: English Nouns, Level: 1, Related to: X,Y,Z..., prerequisites: Q,R,S) however this metadata will also become increasingly large and complex, through learner generated meta data, and updates from our library.

The AI will then be tasked with many processes to perform like:

• 1. The AI will need to create a map of the content, similar to the Khan academy Knowledge map Each learners map may be different as they may all have different content installed on their device, or they may have added custom user generated meta data to a piece of content which affects its position in the map.
• 1. The AI will need to use data it has collected from the user through various interactions, personality tests and their general usage and create dynamic learning plans that constantly need to adapt to the needs, goals and interests of the learner. It will do this by analyzing the personal data of the learner, the meta data of the content, and how the content fits within the relational map of linked content.
• 1. A key focus of this platform is the facilitation of social and collaborative learning. We are using mesh networking technology to link the devices together within the community. The AI’s on the network will communicate with each other —with the permission of the learner— and match up learning pairs or groups to work on problems together. They will also pair Learner X who may have mastered a concept with Learner B who needs help with that concept.

Our current strategy: We intend to start off using almost 100% hard coded recommendation engine type logic then gradually use machine learning to develop these further. The key here is that its not a Big data issue, we calling it a little data issue as we need to gain insights from one users data, on device, and make adjustments locally.

So this is the challenge, how would you structure a solution and what approaches would you use?

Here is some more information about our project: www.dev4x.com