Cutting-Edge AI for Education at the Edge

Jamie Alexandre
Learning Equality
Published in
6 min readMay 14, 2024

I’m excited to share the latest updates on how Learning Equality is leveraging AI to create more equitable learning opportunities for the 2.6 billion people still not on the Internet — including our progress automating the curriculum alignment process, and our plans for getting contextually adapted AI-powered tools directly into the hands of the offline teachers and learners we serve.

lateral bird’s eye view of a boy sitting on a mat on the floor, engaging with a tablet. A workbook by his side.
Kolibri learner at school in Palabek Refugee Settlement, Uganda

The Fallacy of Trickle-Down Tech

Technology is often touted as the great equalizer — that edtech means anyone can now learn for free; that AI lets anyone create art, write code, or automate their work; that social media enables anyone to have voice and influence — but the painful reality is, left to its own devices, tech tends to widen the divides we wish it would bridge. Those with the best infrastructure, the most free time, the most preparation, and whose market needs these tools were designed to address, capitalize on the benefits, while others are left further behind. Reversing this trickle-down trend requires proactively co-designing with and for those furthest from centers of power, whose fundamental needs are not being met due to lack of market incentives.

A core battle for Learning Equality over the past 11 years has been ensuring education technology and open content proactively reach and benefit those without Internet access, and we’ve now helped millions of learners across 220+ countries and territories achieve their educational goals. But with 1/3 of the world still completely offline (and the rate of Internet growth actually slowing), and 1 in 3 children and youth not achieving even basic numeracy and literacy, this work is more important than ever before. Meanwhile, emerging AI tech, though in many cases over-hyped, can also support incredibly powerful use cases for both education and the workforce, but is once again most deeply benefiting those who already have the most.

AI to Automate Curriculum Alignment

Last September, I wrote about how we’re leveraging transformative AI to support curriculum alignment, to streamline the process of organizing content from our vast library of educational materials to various national curricular standards, and ensure offline teachers and learners have easy access to the most relevant materials for their context and needs. After a multi-year consultative process, led in collaboration with UNHCR, I’m excited to share that these tools, which will soon be shared openly as a public good, are now being used in production: A few weeks ago we automatically organized 12,000 learning resources to 2,194 distinct learning objectives in Math, Physics, Chemistry, and Biology, for ten grade levels in the Ugandan curriculum — and these materials are now already being used as part of a UN project, by teachers, students, and out-of-school youth in disconnected communities in Uganda. A process that previously would have taken months of painstaking work costing hundreds of thousands of dollars was reduced to a matter of days, a couple hundred dollars of API credits, and a few thousand dollars of expert review.

After applying our recommender model to each of the 2,194 learning objectives extracted from the Ugandan curriculum, we ended up with 20,698 content recommendations. If a reviewer optimistically took 2 minutes to check each one, that would be 4 months of full time work. Instead, we added a stage leveraging GPT-4 to do a preliminary review of the candidates, eliminating 8,000 lower-quality matches and allowing experts to focus efforts on a final round of spot-checking.
In addition to providing a score of a piece of content’s relevance to a specific learning objective, we had the system also generate an explanation of its relevance (or lack thereof), which could be used to help support expert review, or contextualize the content within an education platform.

Addressing Offline Education Needs with AI at the Edge

We’ve been leveraging AI to benefit offline communities through the ‘back office’ curriculum automation work described above, but it’s now time to get the power of these tools directly into the hands of disconnected communities who don’t currently have access, in thoughtful and needs-driven ways that help to boost equitable learning. We envision three specific areas of need faced by our teachers and learners that these tools could address:

  • Content discovery: allowing learners to express their interests and learning goals through language, then providing them with factual answers that draw from the repository of content loaded onto the offline Kolibri server, and guiding them to the most relevant resources to help them learn more. Teachers will also be able to articulate the learning objectives of their lessons to help discover the most relevant resources to include.
  • Assessments: supporting teachers with drafting novel assessments, targeted to specific learning activities and skills in the learning materials being used on the offline server, for review and use by teachers with their students. Additionally, offline models could provide feedback on assessment responses, identify common misconceptions, and review short and long form writing responses without the need for time-consuming teacher grading.
  • Just-in-time pedagogy: offering teachers recommendations for interventions based on student activity data, suggesting contextually tuned prompting questions to probe for understanding and guide learning, and also recommend additional remedial, supplementary, and extension resources for learners, depending on individual needs.

Cutting-edge commercial AI models are very resource-intensive to run, hosted in very centralized and large-scale cloud infrastructure, and accessed over an Internet connection. Meanwhile, open source models that can in principle be used offline tend to either be smaller with poor performance, or very large as well, requiring expensive and power-hungry hardware. However, we’re starting to see a convergence of two trends, with increasingly promising performance from smaller models (e.g. the recently released Phi-3, which is meaningfully usable even on a Raspberry Pi), and low-cost AI-accelerated hardware being targeted for optimization (e.g. this work to run the state-of-the-art Llama 3 model on the embedded GPU of a $100 Orange Pi, and Beekee’s proof-of-concept and performance testing on the Raspberry Pi).

We can expect these trends to continue, enabling broader and broader usability on lower-cost hardware. In parallel, we need to be building out the integrations, datasets, and fine-tuning processes to ensure that these innovations meaningfully address the educational needs of communities without the Internet. Learning Equality is perfectly situated to take on this challenge, in collaboration with our communities and partners, for several reasons:

  • Kolibri, our open and adaptable set of products designed for offline-first teaching and learning, has broad global reach, seamless mechanisms for offline distribution, and a large diverse library of offline-ready multilingual content.
  • We serve some of the most disconnected learners globally, and have been actively challenging biases in data collection, as part of building a more inclusive and equitable edtech ecosystem. As a result, the data we collect will be able to inform training of AI/ML models that better understand the learning behaviors, constraints, languages, and experiences of learners who are currently under-represented.
  • Our global team has deep expertise in 1) international development and education, 2) optimizing for deployment on diverse low-cost hardware, and 3) developing effective machine learning and AI solutions that address key learning needs. I’m personally very excited to be leading this initiative forward, having been working with machine learning and language models since the early days of my Cognitive Science PhD fifteen years ago.

Help Bolster Innovation in Equitable Edtech

We are excited to announce that our long-time partner and supporter, Endless, has issued a $500K challenge to help Learning Equality pioneer these offline AI tools. We are looking for mission-aligned funders to rise to the challenge and help build the future of equitable AI-powered edtech. To learn more about how to get involved or support, please reach out. Individual donations can also be made through our website.

Sign up to discover human stories that deepen your understanding of the world.

Free

Distraction-free reading. No ads.

Organize your knowledge with lists and highlights.

Tell your story. Find your audience.

Membership

Read member-only stories

Support writers you read most

Earn money for your writing

Listen to audio narrations

Read offline with the Medium app

Published in Learning Equality

Learning Equality is committed to enabling every person in the world to realize their right to a quality education by enabling teaching and learning with technology, without the Internet.

Written by Jamie Alexandre

Executive Director at @LearnEQ, makers of Kolibri: building #edtech to help offline learners and teachers succeed