Pre-school learning with Artificial Intelligence

Education in third world countries is a Catch-22: children have difficulty learning because teachers lack adequate education, the lack of which causes the next generation of teachers to also lack competence. Education is a problem even in western countries: in New York City, 54% of children cannot read at grade level.

The solution is a self-contained system to teach 3-6 year old children reading, writing and maths without human intervention. It comprises puzzle games with letters and numbers in wood and plastic so children memorize the shapes of the letters while having fun. It also has an Artificial Intelligence that teaches children decoding strategies and reading. We have a patent pending for this invention.

At scale, the solution would improve equality of opportunity for millions of children who lack good teachers and deliver on the sustainable development goal of quality education for all.

Children learn reading, writing, and basic math with considerable human involvement, with about one year of contact time with a qualified teacher. This educational model works and has been proven for hundreds of years, but it has two problems.

First, it perpetuates social inequality within western countries, as high-quality teachers prefer to teach in high-quality schools attended by children from privileged families, and children from disadvantaged backgrounds learn from lower-rated teachers. Children from privileged backgrounds gain a considerable head start that persists through adult life. Low early literacy decreases wages and educational attainment (see Elango et al., 2015) and increases health care costs (see Concordia University).

Second, it perpetuates social inequality between countries, as emerging countries lack qualified teachers to provide a good education (Deon Filmer found that "fewer than 3 percent of teachers in Mozambique, Nigeria, and Togo can competently grade homework based on the curriculum they teach"), and without a good education the next generation of teachers will also lack competency.

Early childhood interventions have very high returns of 7-10% per year (Elango et al. 2015). The flip side of this result is that children who don't receive good interventions will lag significantly behind.

We are working with pre-school children in western countries committed to early childhood education, such as France, and in emerging countries, such as Mozambique. France has made kindergarten mandatory from age 3 to combat the early literacy gap and we see our solution as part of that mission. We target public schools in under-served and disadvantaged areas that teachers with higher ratings would rather avoid. We use the proceeds from these areas to subsidise operations in emerging countries, in particular Mozambique, where the founder has visited and worked closely with a pre-school in the province district of Inhambane.

The solution is a self-contained system to teach 3-6 year old children reading, writing and maths without human intervention. It consists of a physical aspect and a software aspect. The physical aspect is similar to the game of shapes with triangles, squares, and circles where children have to fit the right letter in the right slot; the difference is that it consists of a wooden board with letters that spell the child's name (their favorite word), children play the puzzle of finding the letters that go in the right slots, and the plastic letters on wood form a nice result (unlike the game, the letters stay and do not fall through). It also has individual letters for children to form any word, numbers to learn to count, and math symbols to learn arithmetic. Children learn the shapes of letters from repetition in the sheer amount of times they play with them.

The software aspect consists of an artificial intelligence that teaches children reading, writing, and maths. It is embedded in a server box with a camera, a screen, a microphone, speakers, and buttons. Children can speak into the microphone and a speech recognition system recognises the word and shows its spelling on the screen. Children can learn new words prompted by their curiosity and at their own pace.

The software can also pronounce any word that children form with the letter blocks. Upon forming a new word, children place the word under the camera, press a button, and the software recognizes the word with Optical Character Recognition and pronounces it with speech synthesis, so children learn to explore the language at their own pace. Depending on the ability and progress of the child, the software teaches decoding strategies by sounding the relevant spelling rule, e.g. "EA" often makes the [i] sound (the closed, front, unrounded vowel) as in "TEA" and "EAT", but sometimes it makes another sound (the open-mid, front, unrounded vowel) as in "tearing apart". These strategies are especially important for orthographically deep languages like English, Portuguese, and French.

We did extensive testing and the physical part is now feature complete. We designed a typeface that minimises frustrations for children, e.g. children expect the E to fit into the slot either right side up or upside down. We have filed a patent for this invention.

The toy innovates by merging a physical toy appropriate for young children with the flexibility of software to teach them literacy. Many tablet applications that help with literacy are all-digital and unsuitable for very young children.

The dual physical and virtual aspects allow children to play with a physical toy that evolves with them. For example, the toy teaches decoding strategies to children appropriate for their age and previous experience, so children are always learning new strategies. Children can explore the language, e.g. they can use the wooden and plastic blocks to form a word that does not exist, such as "wugs", press a button on the server box, and an artificial intelligence synthesizes the pronunciation of that word from having digested a phonetic dictionary of the language. To play with words, the artificial intelligence proposes new words from the letters: for example, if children form "nod", the system proposes "node" with a new letter or "done" with a new letter and a rearrangement of letters in increasing complexity. To teach maths, children can form equations such as "1+1=2" and the system provides feedback on whether it is correct.

We made many improvements based on children's feedback, e.g. that some letters like the S need to be symmetric to avoid frustration. Our extensive testing leads us to believe that the physical part of the toy is now feature complete and we have applied for a patent.

The solution uses software and technology at every stage of the product. For the hardware part, the choice of letters is optimized with linear programming to the most common words in children's books of the target language. From this choice of optimal character combinations, software automatically generates the design files for Computer Assisted Manufacturing. Fabrication is digital with computer-numerical control (CNC) laser-cutter and router. The hardware is all streamlined to keep costs low. At scale, the molds for the pieces will be machined to drive costs even lower and save on material.

The software part is where we most excel at technology and we will continuously improve. Computer vision methods (such as edge detection, histogram equalization, and contour simplification) and machine learning methods (nearest neighbors and neural networks) serve for optical character recognition. Neural networks serve for speech recognition, which are pre-trained on massive datasets and that we fine-tune for children's voices on a small dataset. Speech synthesis allows the software to communicate with the students. Neural networks also serve to digest a full dictionary with words and phonetic transcriptions and synthesize the pronunciation of new words. This part has fixed costs and can be deployed at zero marginal cost, which also means that any improvement can affect every single user with updates.

We are currently at the prototype stage. While it may sound early to apply for MIT Solve, we see our solution and mission to be so aligned with the challenge of early childhood education and the sustainable development goal that we decided it was too good to miss. Therefore, our belief that the solution addresses the problem is based on preliminary results from our experience.

The main results stem from experience with children who started playing with the device when they were two years old. After one year they had taught themselves to read and write their own names. After two years, they can read and write all names in the extended family. They have learnt to recognise all letters of the alphabet, they know the correspondence of case between capital and lowercase letters, they can read and write all names in the extended family, they recognize all numbers, and they haven't even started compulsory education. After two and a half years, they still enjoy playing with the toy.

We believe that children could use this system to learn fast and learn well because they are learning at their own pace, from their own natural curiosity, and with a system that evolves with them. By repeatedly playing with the letters, they memorize their shape and are naturally inclined to copy them with a pen, teaching themselves to write. We have got a lot of interest from parents who like the toy and want their children to learn while having fun.

We are currently serving 10 children. In one year, we will be serving 150 children for the pilot with 5 kindergartens. In five years, we will be serving at least 20% of the children in one geographic area, where 20% is a target for the schools from disadvantaged backgrounds and the geographic area could be the province of Québec, the New York City public school system, or the entire French public kindergarten system.

Over the next year we want to pilot a project in 5 schools with 150 children total, incorporate the lessons from the pilot into the randomised control trial (performed by external academic partners) over two years to gauge the effectiveness of the solution, and use the results from the trial to sell to the large organisations we want to reach in 5 years.

Over the next year, the main barriers are the market fit and finding a set of ideal clients to make the business sustainable; and recruitment of a larger team to scale the business (which has been a one-person team for two years). Over the next five years, the main barrier is being ready for the randomised control trial because early childhood education takes several years to show results, so the trial must be ready by year 2 if we want results by the start of year 5.

Regarding the ideal client, I am using a business coach with a track record in making early startups successful and making million-dollar sales. Regarding team recruitment, I feel confident that exposure at MIT Solve will provide the opportunities to meet talented people committed to the same vision of quality education for all. Regarding the randomised control trial, I have a network of researchers from my previous career in academia who specialise in early childhood education and a list of several grant bodies with programs to fund rigorous evaluation of innovative solutions in early childhood education, such as the Education Endowment Foundation in the UK, which also offers support in taking positive approaches to the national level.

One person (founder and CEO), working on this solution part-time at the moment (alongside a part-time job) and working full-time starting in September.

I hold a Bachelor of Science in Computer Science from Ecole Polytechnique (France) and a PhD in Economics from Columbia University. I did a post-doc in Economics at the University of Cambridge focused on early childhood education and another post-doc in software engineering and artificial intelligence at The Alan Turing Institute. I am passionate about children's education and have invented several toys, devices, and games. I also enjoy digital fabrication with graphic design and computer-assisted manufacturing. I am committed to open-source software (see my Github repositories at github.com/miguelmorin and my contributions on StackExchange @mmorin) and decided to file a patent for the sustainability of the business. I am also passionate about languages (speaking fluently Portuguese, French, English, Spanish, and having passing knowledge of German).

We currently use the services of Microsoft Cognitive Services for speech recognition and speech synthesis. We will develop our own system that can work without an internet connection. We work with them with automated queries to their Application Programming Interface.

In the early years, the key customers are private foundations in western countries that manage their parents' corporate social responsibility budget and target early childhood education and equality of opportunity and fund service provision in the communities of the parent companies. The unique selling proposition is innovation and high social returns regarding educational achievement and social inequality. I have identified 24 such foundations.

Later on, the key customers are governments sensitive to early childhood education and social inequality, such as France and the UK. The unique selling proposition is high-quality and low-cost early childhood education adapted to limited budgets and can be deployed where high-quality teachers would rather not go.

We have personal savings to cover two years of operation and are also applying for grants for the early years. Financial sustainability afterwards will be from customer revenue, which is the best way to stay in business.

I have spent 10 years in academia, where thinking is narrowly defined, delimited with specific steps, and mostly negative in the sense that people shoot down ideas to see which ones survive. I would benefit from interacting with creative thinkers, optimists, and courageous creators who are not afraid to take on challenges. I have made big leaps in my solution thanks to feedback from other people and I expect I will benefit even more from this world-class selection of talent in pivoting my business model or improving the product. We would also benefit from the exposure to recruit skilled and committed people into our team when we are ready to grow, and possibly also to find partners from the Solve network that share the vision of the sustainable goal 4, quality education for all. And we would also benefit from the financial stipend.

I would like to partner with organizations that have expertise in running early childhood interventions, either in western or emerging countries.

We would like to develop our own system for speech recognition and speech synthesis that works without an internet connection and that targets languages in emerging countries, such as Swahili.

As we already use AI in education, we will use the prize to set operations in the United States, where we are not currently operating.