5 level taxonomy of AI in learning (with real examples)

AI fallacy 1: dystopia
Let’s not be misled by dystopian Hollywood visions of AI. Movies like Her, Ex Machina and Chappie are fiction and this is about fact. Robots have certainly had an impact in manufacturing (as did machines in agriculture when labour moved from fields to factories), where their speed, precision and ability to deliver 24/7 have led to massive increases in productivity. The cost has been the elimination of dull, monotonous and repetitive jobs. But AI is a broad and complex area of endeavour, only one of which is robotics.
AI fallacy 2: mimics the brain

Neither should we see AI as simply analogous with the human brain. This is another AI fallacy. We didn’t succeed in the airline business by aping birds, nor did we make much progress in going faster by copying the legs of a cheetah – we invented the wheel. So it is with AI. It’s about doing things well, more consistently, faster, more accurately than the human brain. Our brains have several drawbacks when it comes to some real world tasks. It likes to spend one third of its time asleep, the other third in leisure. It is also full of biases, gets tired, inattentive, has emotional swings, even suffers from mental illness.

Similarly, in the world of learning, AI is not about dystopian fantasies or aping teachers. AI is already being used by almost every learner on the planet, through that algorithmic tool Google. It is already being used in predictive analytics and already being used in adaptive learning.

5 Level taxonomy of AI in learning
To untangle some of the complexity I propose a five level taxonomy for AI in learning. My taxonomy is similar to the five level taxonomy developed for automated vehicles, where the driver is in complete and sole control of a vehicle, with only some interal algorithmic fucntions obvious on the dashboard, through assistive power steering, predictive satnav tech, therough degrees of autonomy, to full self-driving automation. At the top level, vehicles are designed to perform all safety-critical driving functions and can safely operate without any driver intervention.

Level 1  Tech

Level 2  Assistive

Level 3  Analytic

Level 4  Hybrid

Level 5  Autonomous

Level 1  Tech

You’re reading this from a network, using software, on a device, all of which rely fundamentally on algorithms. These include; Public Key Cryptograph, Error Correcting Codes, Pattern Recognition, Database use and Data Compression – to name but a few. With data compression, we when we use files, they are compressed for transmission, decompressed for use. Lossless and lossy compression and decompression magically squeeze big files into little files for transfer.

These, and many other algorithms, enable the tech to work and shape the software and online behaviours of people when they are online. These algorithms really are works of art that have been designed, tweaked and finessed in response to experiment with real hardware and users. They work because they’ve been proven to work in the real world. Of course, what’s seen as an algorithm is likely to be multiple algorithms with all sorts of fixes and tricks. These ‘tricks’ of the trade, such as checksum, prepare then commit, random surfer, hyperlink, leave it out, nearest neighbour, repetition, shorter symbol, pinpoint, same as earlier, padlock,, these make algorithms really sing. Every time you go online all files you use, audio you hear, images and videos you watch, are only possible because of an array of compression algorithms. These are so deeply embedded in the systems we use they are all but invisible. The personal computer you use is essentially a personal assistant that helps you on your learning journey. With mobile you now have a PA in your pocket. These are examples of AI and algorithms deeply embedded in the technology and tools.

Level 2  Assistive

Google was a massive pedagogic shift, giving instant access to a vast amount of human knowledge teaching and learning resources. Yet Google is still simply an algorithmic service that finds and sorts data. Every time you enter a letter into that letterbox it brings huge algorithmic power to bear on trying to find what you personally are looking for. Search Engine Indexing is like finding needles in the world’s biggest haystack. Search for something on the web and you’re ‘indexing’ billions of documents and images. Not a trivial task and it needs smart algorithms to do it at all in a tiny fraction of a second. Then there’s Pagerank, now superseded, the technology that made Google one of the biggest companies in the world. Google has moved on, or at least greatly refined, the original algorithm(s), nevertheless, the multiple algorithms that rank results when you search are very smart.

Other forms of assistive, algorithmic power in learning include; unique typing and facial recognition in online assessment. Pattern Recognition is just one species of algorithms used in learning. Learning from large data sets in translation, identifying meaning in speech recognition – pattern matching plucks out meaning from data. Mobile devices especially need to use these algorithms when you type on virtual keyboards or use handwriting software. Facial and typing recognition are now being used to authenticate learners in online assessment.
A nice example of assitive AI in learning is PhotoMaths, which uses the mobile phone camera to 'read' maths problems and not only provide the answer, but break down the steps to that answer. Algorithms are therefore increasingly used to directly assist learners in the process of learning.

Level 3  Analytic

Using algorithmic power to analyse student, course, admission or other forms of educational data, is now commonplace. Here, an institution can mine its own, and other, data to make decisions about what it should do in the future. This could be increasing levels of attainment, identifying weaknesses in courses, lowering student dropout and so on.

Beyond the institution, on MOOCs, for example EdX have identified useful pedagogic techniques, such as keeping video at 6 minutes or less, based on an analysis of aggregated data across many courses and many thousands of students. Smart algorithmic analysis an also identify weak spots in courses, such as ambiguous or too difficult questions.

Level 4  Hybrid

Technology enhanced teaching where algorithmic power is applied to the tasks of teaching and learning. Here the AI powered system works in tandem with the teacher to deliver content, monitor progress and work with the system to improve outcomes.

A good example is automated essay marking, where the system is trained using a large number of professionally marked essays. These marking behaviours are then used to mark other student essays. For more detail see Automated essay marking - kick-ass assessment.

Another example would be spaced practice tools, that often use algorithms such as SuperMemo, to determine the pattern and frequency of spaced practice events. See an example here as used by real students.

However, the most common use is in adaptive learning systems, where the software uses student and aggregated student data to guide the learner, in a personalised fashion, through a course or learning experience. This is still in the context of a human teacher, who uses the system to deliver learning but also as a tool to identify progress among large numbers of students and take appropriate action. We are educating everyone uniquely but it is still technology enhanced teaching. A good example is CogBooks. This is where we are at the moment with AI in learning. Our evidence from courses at ASU suggest that good teachers plus good adaptive learning produces optimum results.

Level 5  Autonomous

Autonomous tutoring is the application of AI to the issue of teaching without the participation, even intervention, of a teacher, lecturer or trainer. The aim is to provide scalable, personalised solutions to many thousand, if not millions of learners, at very low cost. The software needs to be able to deliver personalised content based on user data and behaviour, as well as assess. In some cases autonomy can rise to a level where the system learns how to deliver better learning experiences, on its own, to produce self-improvement, through machine learning. There are already online systems that attempt to do this, such as Duolingo, used by over one hundred million learners and other platforms are on their way to performing at this level.
At this level, one could argue that the concept of teaching collapses. There is only learning. In the same way that Google collapsed the idea of the person looking through shelves on libraries or card indexes tos earch and find information, autonomous AI will disintermediate teaching.

Other dimensions

This taxonomy looks at AI from the educators perspective and works back from the learning task. Another perspective is the different types of AI that can be applied in learning. Looking at our taxonomy again, one can identify algorithmic power that delivers, technical functionality, speed/accuracy of learning task, speed/efficacy of predictive analytics, algorithms that embody learning theory, Natural Language Processing, genetic algorithms, neural networks, machine learning and many other species of algorithm.

Within this there’s a plethora of different techniques using data mining, cluster theory, semantic analysis, probability theory and decision making that takes things down to the next level of analysis. This, in my view is too reductive. Yet this is where the real work is being done. This is very much a field where real progress is being made at a blistering pace, fuelled by massive amounts of data from the internet. But this approach to taxonomy is of little use to professional educators who want to understand and apply this technology in real learning. Contexts.

Conclusion

AI in learning is not without its problems, in terms of privacy, false positives, errors, over-learning and potential unemployment. Some of these problems can be overcome through progress in the maths and design, others lie on the regulatory, cultural and political sphere. But given the fact that productivity has stalled in education and costs still rising, this seems like a sensible way forward. If we can deliver scalable technology, assistance, analysis, learning and teaching, at a much lower costs than at present, we will be solving one of the great problems of our age. Macines helped move us on through the industrial revolution, where machines replaced manula labour. They will now move us on replacing some forms of mental work. When a famous economist standing at a huge building site asked of a government official “Why are all these people digging with shovels?” the official proudly said “It’s our jobs programme”. The economist replied “So then, why not give the workers spoons instead of shovels?” We are still, in education, using spoons to educate.