Sunday, July 03, 2016

10 important things AI teaches us about ‘learning’

Alvin Toffler, who died this week, said, The illiterate of the 21st century will not be those who cannot read and write, but those who cannot learn, unlearn, and relearn.” But he had a better quote, “If you don't have a strategy, you're part of someone else's strategy.” There’s one species of this argument that really matters in the learning game - if you don’t have an AI strategy, you’ll be part of someone else’s strategy.
Most pedagogic change now comes through the use of technology. The internet is a huge Darwinian machine that selects ‘fit for purpose’ services, which millions, sometimes billions, want and use. As a result, in the learning game, we have seen more pedagogic change in the last 20 years, than the last 2000 years. This process is accelerating.
There are some new kids on the learning block, like blockchain. But at a deeper level there’s something far more significant that’s happening; AI is re-shaping the learning landscape. Intriguingly, AI now draws its inspiration from evidence of how we actually learn. This should be a wake up call for those of us who work in the field. The AI community now shows us what works, practically, in learning theory, and what doesn’t. They do this by building ‘learners’, software that learns. Pedagogic change no longer comes from educational research (not sure that it ever did), it comes from insights in cognitive science and, increasingly, through this form of technological innovation. AI is the latest manifestation of digital pedagogy and the one that is now giving us confirmation about what works and doesn’t work.
Learning embedded in AI
We’ve recently seen some super successes in AI, with Deepmind’s spectacular win against one of the World's greatest GO players, the Todai project passing the Tokyo University entrance exam. Deepmind used a layered neural network, with an executive layer, initially trained on 30m human moves, which then played itself using a trial-and-error process (reinforcement learning), with the huge processing power of the Google Cloud Platform – and it won.
This is where it gets fascinating for those of us interested in learning theory. Cognitive scientists, and cross-discipline minds, like Demis Hassabis, have taken principles from cognitive learning theory, such as clustering, attention, learning by doing, reinforcement learning, chunking and practice, embodied these in AI, and are using them to great effect in getting software to learn and problem solve. These learning machines are showing us the way. So what can WE learn from this?
1. Search
Access to knowledge and skills has long been a problem in education and training. The traditional model has been slow, scarce and costly supply through expensive institutions such as Universities and libraries, with rising costs and debt. Then along came Google, a great pedagogic leap, and reduced that time and cost to almost nothing. This has been a huge pedagogic leap, one that completely reshaped the learning landscape at all levels. It is all down to AI. Google is nothing but AI.
2. Feedback
We know that improved and detailed, personal feedback accelerates learning. Yet traditional teaching, especially in the classroom and lectures, make this very difficult. Using data to dynamically adapt, in real time, what is taught next, is a common technique in AI. Much of what you see online is determined by algorithms that constantly monitor your needs. Software (AI) driven feedback is the only way to provide such detailed feedback, personally, on scale. AI thrives on new data to update what it thinks it knows, whether from the individual learner or aggregated learners. This is how Bayes and dozens of other species of AI algorithms work. We need to recognise that the brain has exactly the same needs. Everyone has unique learning needs and everyone need to be educated uniquely.
3. Less is more
The principle of clustering, indeed most refinement of algorithms, reduces what has to be learnt to a minimum, looking for optimal ways forward, while retaining efficacy. This is an important ‘less is more’ principle that is all too often ignored in real world teaching and learning. We still teach at too general a level, for example, teaching how to write essays by repeated essay writing, rather than the more detailed components of good writing and analysis. For a complete breakdown of this error see this excellent video by Daisy Christodoulou. AI has applied and developed a battery of mathematical techniques to optimize learning. They select, reduce, optimize, judge and create recommendations. It has 2500 years of philosophy, logic, probability theory and mathematics behind it and now that we have an abundance of data and cheap computers, is bearing astonishing and exotic fruit.
4. Chunking
We have known about this for decades. Chunking, meaningfully, accelerates learning. AI does this often with data structures, pre-processing and choice of efficient algorithms. Sorting algorithms are a great example of chunking (binary sort) for efficieny. Skills acquisition is not about just, say playing golf, but practising how to putt, drive etc. No golfer becomes great by simply playing golf – they chunk down and practice. AI has shown that, to learn effectively, you chunk problems down into their constituent parts, practice those, then build up your skills. AI is lots of little skills that add up to something big – like the self-driving car. Who saw that coming? Teaching general skills is still all too common, especially in schools, colleges and universities. But it’s operating at the wrong level. AI has taught us that we need to focus on the detail.
5. Reinforcement learning
Another insight from learning theory, that the AI folk have picked up on, is that people learn by DOING things. Children don’t learn much by sitting and listening. They do stuff. This reinforcement learning, the idea that every state has a value, beyond the simple binary ‘win’ or ‘lose’ positions, is powerful. Reinforcement learners (software) rehearse learning a huge number of times, trying and trying again, sometimes the best way, sometimes randomly. It’s a turbo-charged learner. So having learnt what to do from human data, it plays itself, an enormous number of times in a short period of time, to get super-smart. This is how Deepmind beat the GO champion. They have been hugely successful in all sorts of AI, real world tasks. It tries all sorts of habits but selectively chooses those that are successful. It learns how to learn. This effortful learning, learn by doing theory, is at the heart of AI.
6. Deliberate practice
AI has embodied the idea of ‘deliberate practice’, from Ericsson, and built algorithms and methods of propagation that do exactly that. They practice with intent, that intent being improved performance. This is what backpropagation in neural networks and many other machine learning approaches do – they automate and optimize deliberate practice and improvement. They embody what education and training has ignored for too long – deliberate practice.
7. Spaced-practice
Spaced-practice tools are largely driven by algorithms that deliver the pacing, interleaving and load balance for spaced practice. We have known since Ebbinghaus, since 1885, that learning suffers from massive forgetting. We also know that the solution is deliberate, spaced-practice. This can be effected online with smart algorithms that determine how this should be delivered. They do what no teacher can do, identify what needs to be reinforced, how it needs to be reinforced (interleaved etc) and when it needs to be delivered (load balanced etc), related to that individual’s needs. It is all down to AI.
8. Data driven
We are no longer data poor, we are data rich. It is no accident that slow burning AI suddenly had its Cambrian explosion, with thousands of practical examples, from speech recognition to self-driving cars, breakthroughs that are reshaping global industries. The ability to manage, read and interpret this data gives you the radar you need to keep ahead of the game. That’s exactly what AI does applying logic, probability and computer power to the problem of prediction. It has been fashionable to ignore ‘knowledge’ as just ‘data’ in education, but AI has shown that knowledge (data) really does matter. It is an integral part of learning, not something to be abandoned or tritely classed as ‘rote learning’. It is, in fact, what enables deep learning.
9. Socratic learning
Socrates learning theory is often called a ‘theory of ignorance’. It was a process of excising what you think you know, stripping things back until the real knowledge was exposed. He taught us humility in learning. This is also true of AI. It knows what it knows but also knows what the probabilities are in its outputs. In this sense it is free from the cognitive biases, even gender, race and socio-economic biases that teachers, as they are human, almost always have. We have a chance here, to both recognize our limitations and embrace technology enhanced teaching. A little Socratic humility, through the use of AI, recognizing where we are good, and not so good, in teaching (and learning), could go a long way.
10. Learning to learn
An interesting insight comes from what AI folk call ‘learners’. These are machine learning algorithms, that learn themselves. This fundamental point, that the new landscape is not the old one of ‘human teachers’ and ‘human learners’, but also one of’ machine teachers’, and importantly, ‘machine learners’. These new entities cope with complexity; time complexity, space complexity and errors. The machine learners learn to learn and now learn how to teach to learn. The automatons automate automation. This has profound implications for productivity, employment and politics. It will, in time, become an existential issue, first for the professions, even for our species.

AI is all about learning. It is software that learns. It is also software that can create good learning content. We have to pay attention to what is happening here. AI tells us that learning is a process, one that can be unpacked and copied - reverse engineered. They are achieving things that were unimaginable just a few years ago. They are learning about learning by creating effective learners. The fact that AI is having so much success by following the lessons that cognitive science has to teach, is surely a wake up call for learning theorists, especially those still stuck in foggy world of social constructivism.

No comments: