Learning about learning via virtual students

AI is the hot area for start-ups right now following the $500m acquisition of DeepMind by Google and the reported $150m acquisition of Magic Pony by Twitter. Valuations for start-ups in this area are shooting up and the ability to raise money for such ventures seems endless. It’s 25 years since I did my PhD in Robotics and AI and the momentum for innovation in this field is rapidly changing.

In education, we’re seeing AI being used to optimise both learning and teaching.  

Where a vast majority of edtech tools are failing to make a fast enough difference for the children who need help the most and the burnt-out teachers who are striving to support them, AI could be the key to solving one of the biggest social conundrums that has foiled governments and academics for years. 

Will AI democratise access to quality education? And can it remedy some of the wider issues beleaguering the teaching profession today – from recruitment and retention to engagement and work life balance?

I believe so, but we are at the start of a long journey and efficacy of education technology largely lives and dies at the hands of politics, resourcing and lack of funding rather than the quality of the products themselves.

Will AI democratise access to quality education? And can it remedy some of the wider issues beleaguering the teaching profession today – from recruitment and retention to engagement and work life balance?

From first-hand experience and through observing the work of other innovators implementing technologies in particular schools, the impact can be meaningful, tangible and – to some degree – life-changing.  The potential for positive change in education and the repercussions it will have socially, democratising potential and opportunity, are truly promising.

The role of AI in edtech innovation is particularly exciting and, it is for this reason, I have become involved in an initiative that is seeing the world’s first ever ‘virtual student’ being built.  

Building a virtual student goes right to the heart of AI research which, in part, asks a fundamental question that must be crucially addressed by all edtech developers: how do humans learn and can we emulate this in software?

When delivering AI-based projects, you need to have a large enough data set in order to train and validate before running real world tests to see whether the algorithm works in practice. Collecting a large data set takes time and is compounded in the education space by significant privacy issues around children’s data. However, data created from a virtual student, a software program that simulates a real student, can be freely shared. Moreover, running tests in the real world on real students is problematic as you may be impacting real students’ progress and it may also take many weeks, months or even years to find out how well your software works. However, with a virtual learner you can simulate a year’s learning in a few seconds.

Building a virtual student goes right to the heart of AI research which, in part, asks a fundamental question that must be crucially addressed by all edtech developers: how do humans learn and can we emulate this in software?

Of course, it might take many years before we have anything even approaching how a real student might learn, but the initiative will lead to some interesting insights that may help us to address this bigger question.

Adaptive algorithms in their own right are an exciting area of progress in AI. Professor Hermann Ebbinghaus discovered what has since become known as the “forgetting curve” in 1885, but it was only in recent years that UK companies like Memrise, Duolingo and Lingvist then developed these learnings into algorithms that have made it into mainstream learning.

Ebbinghaus discovered, through empirical observation, that if you learn a fact, then your likelihood of remembering it over time decreases exponentially.  A further finding was that if you are asked to recall that fact just at the point of forgetting it, then the probability of you remembering it is pushed back to 100%, but, more importantly, the slope of the forgetting curve decreases and you will be able to recall the fact for longer. Indeed if you relearn a fact five to seven times you are likely to remember it forever. 

I am amazed that children today aren’t taught about the forgetting curve at school and that schools don’t use spaced repetition software to help children retain information. My son has just completed his Year 10 and gone through 15 modules in maths. At the end of each module he took a test. It’s now the end of the year and he must revise all 15 modules for an exam in two weeks’ time. It would be far better if schools were using spaced repetition software to optimise the long-term retention of, in this case, mathematical techniques. But you can see how this can be applied to any discipline.

If you relearn a fact five to seven times you are likely to remember it forever. 

As we progress to develop our virtual student, one of the key challenges we face will be in determining the appropriate metric for optimised learning. For example: given X number of minutes of study per day, what algorithm optimises the number of facts that can be recalled after Y days? And what of facts retention – given X number of facts to be learnt and Y minutes of study per day, what algorithm minimises the number of elapsed days to reach 100% recall?

Of course I’m a firm advocate of edtech, I believe in the possibilities of AI, yet I am always reminded of the advice my PhD supervisor gave me: AI is always the second best solution to solving any problem.

We wait with bated breath to see if modern AI approaches can outperform the best-known algorithms that have been developed over the last 50 years. And if it does, we can expect to see a radical shift in the way we learn and teach through edtech for the better.

Charles Wiles is CEO of edtech venture Zzish, a software company that specialises in transforming all e-learning apps into classroom-ready tools and gives real time analytical insight on student and class performance.