“ The Science of Learning”
1: The Backdrop
2: Development of the Sciences
3: Educational practice ignores research
4: Bloom’s 2 sigma problem
5: What machine learning informs us about human learning?
6: The current state of assessment
7: AInEd Supremacy
1: The Backdrop
I have just been hearing all the noise about the notifications and preparations for the CBSE and other Board examinations. Students aged 16 to 18 years, would be wasting their next 6 months in a completely detrimental activity. Some will also be spending substantial time and money with coaching Institutes doing more of the same.
In a recent (2018) book “ The fourth Education Revolution”, its author Sir Anthony Seldon lists (pages 165 and 197) the negative attributes of these monolithic examinations and celebrates their imminent death. This is not the only example of the divergence of educational practice from what research says.
To begin with, the learning pyramid ( https://www.educationcorner.com/the-learning-pyramid.html) suggests that most students only remember about 5% of what they hear in lectures, 10% of what they read from textbooks, 50% of what they discuss but retain nearly 90% of what they learn through teaching others. Yet our dominant form of teaching is lectures. The CBSE prescribes the mandatory number of these mostly ineffective lectures, which must be attended before being allowed to take examinations. Shouldn’t we be appreciating more those who can learn without attending lectures.
It shouldn’t surprise us that most School leavers are not ready either for the world or for further education, and are ready to be “the useless class” of the 21st Century described by Yuval Noah Harari ( https://ideas.ted.com/the-rise-of-the-useless-class/).
The practice of medicine progresses because the outcome of medical research is applied to medical practice. We saw last year that medical science made a big jump in understanding and then making products with RNA vaccines for the first time in record time while dealing with the Coronavirus vaccine. But in the corresponding urgent matter of making our youth future ready and relevant, the policy making is driven in a leisurely mode, with no reference to the Science of learning, sometimes in complete opposition to what learning science tells us.
Stanford University Professor Savas Dimopolous says “ What differentiates scientists is an intrinsic ability to discern a good idea, a beautiful idea, worth spending time on and most importantly what is a problem that is sufficiently interesting and difficult that it hasn’t been solved yet and now is the time for solving it “. The Science of Learning is such an emerging field.
The Science of Learning is building on the advances in neurosciences and the availability of large amounts of data, which drove learning analytics and educational data mining. Now with Machine learning and deep learning helping us understand how machines learn, there can be new research and applications in the science of learning. Just as advances in Physiology/ medicine as well in Physics and Chemistry spurred the development of the Pharmaceutical Industry, we will see the development of a whole new educational ecosystem in the coming decades. An educational diagnostics industry will replace the erstwhile School leaving Board and end of year University examinations. This has been very well articulated in Anthony Seldon’s book “ The fourth Industrial Revolution” at p.165 and 197 referred to earlier in this piece.
There are connections between, learning, computing and brain science which are well illustrated in this TED talk by Jeff Hawkins:
Another closely connected idea to human learning is consciousness. Roger Penrose on Quantum Physics of Consciousness : https://youtu.be/43vuOpJY46s
About a hundred years ago, Quantum Mechanics developed with a number of persons attempting to overcome the limitations of classical Physics, and now we are witnessing the 2nd Quantum Revolution. Similarly developments in the Sciencr of Learning will drive the 4th Education Revolution.
2: Development of the Sciences
In prehistoric times, knowledge and technique were passed from generation to generation in an oral tradition. For instance, the domestication of maize for agriculture has been dated to about 9,000 years ago in southern Mexico, before the development of writing systems. Similarly, archaeological evidence indicates the development of astronomical knowledge in preliterate societies.. The development of writing enabled humans to store and communicate knowledge across generations with much greater accuracy.
If we fast forward to the early 19th century, this was a period that shaped science. Major breakthroughs came in biology, especially in Darwin’s theory of evolution as well as physics (electromagnetism), mathematics (non-Euclidean geometry, group theory) and chemistry (organic chemistry).
A paradigm shift, a concept identified by the American physicist and philosopher Thomas Kuhn, is a fundamental change in the basic concepts and experimental practices of a scientific discipline. Even though Kuhn restricted the use of the term to the natural sciences, the concept of a paradigm shift has also been used in numerous non-scientific contexts to describe a profound change in a fundamental model or perception of events.
Kuhn presented his notion of a paradigm shift in his influential book “ The structure of Scientific revolutions” published in 1962.
Kuhn contrasts paradigm shifts, which characterize a scientific revolution to the activity of normal science,which he describes as scientific work done within a prevailing framework or paradigm. Paradigm shifts arise when the dominant paradigm under which normal science operates is rendered incompatible with new phenomena, facilitating the adoption of a new theory or paradigm. This is what happened in the first 3 decades of the 20th Century with the development of Quantum Mechanics. A hundred years later we are witnessing the arrival of Quantum Technologies.
In my view “ the science of learning” is right now in the pre-paradigm shift stage, and the coming 3rd decade of the 21st Century will see a new phase of “the science of learning”, much like the Copernican revolution. Or like Quantum Mechanics and Classical Physics. The current practice of education is driven by intuitively obvious but incorrect widely held beliefs, and it is now time to upgrade it by adopting research findings and evidence driven practices.
3: Educational practice ignores research
I have made this comparison often. Till about a hundred years back, medical practice was like the present education system, disconnected from the science of Physiology. In his book, “ The Youngest Science” Thomas Lewis shares how he first learned about medicine by watching his father practice in an era when doctors comforted rather than healed. Looking back upon his experiences as a medical student, young doctor, and senior researcher, Thomas notes that medicine is now rich in possibility and promise. The medical fraternity adopted both science and technology, and integrated it so well that when the Cornavirus pandemic came, most people were surprised that the medical establishment had no treatment for it. This amazing progress in medicine has been achieved because the outcomes of research in medicine and Physiology are applied to medical practice.
In the practice of education, on the other hand, the knowledge from research and experiments is not only ignored, educational practices and policies including the NEP 2020, do just the opposite.
The “learning pyramid”, sometimes referred to as the “cone of learning”, developed by the National Training Laboratory, suggests that most students only remember about 10% of what they read from textbooks, but retain nearly 90% of what they learn through teaching others. The Learning Pyramid model suggests that some methods of study are more effective than others and that varying study methods will lead to deeper learning and longer-term retention.
The Learning Pyramid suggests that “Lecture” is one of the most ineffective methods for learning and retaining information. Lecture is a passive form of learning where you simply sit back and listen to information being spoon fed to you by your teacher or professor. Attempting to acquire information and gain understanding only through lectures is not the most effective way of learning. Its effectiveness is only 5%. Audio-visuals have an effectiveness of 20%, demonstrations30%, discussions 50% and practice at doing is 70% effective.
But teaching others is 90% effective. This is what the Physics Nobel
Laureate Richard Feynman also advised. He recommended the following steps:
1: Write the name of the concept at the top of a blank piece of paper.
2: Write down an explanation of the concept on the page. Use plain English. Pretend you are teaching it to someone else (e.g a new student). This should highlight what you understand, but more importantly pinpoint what you don’t quite know.
3: Review what you have pinpointed you don’t know. Go back to the source material, re-read, and re-learn it. Repeat Step 2.
4: If you are using overly wordy or confusing language (or simply paraphrasing the source material) try again so you filter the content. Simplify your language, and where possible use simple analogy.
But the present education systems emphasises attending lectures, to the extent that at both School and at higher education levels there is a mandatory requirement to attend 75% of these rather ineffective lectures, before they can take the terminal examinations. The other more effective ways of learning are seldom used. Even the duration of the lectures, is several multiples of the attention spans of the learners, as indicated by research. There is sufficient educational research to suggest that ’chunking’ of learning is more effective than long duration lectures.
Content should therefore be in the form of ‘learning objects’ that are re-usable and available as a repository of re-usable learning objects. Preference should be to nano-learning objects that can be transacted or perused in 1,2,5 or 10 minutes, and any learning episode may be built by putting together these learning objects according to the pedagogical requirements. Each learning object will have meta-data to allow automation in their search, retrieval and combination.
4: Bloom’s 2 sigma problem:
In the early 1980s, renowned educational psychologist Benjamin Bloom made a breakthrough discovery. Examining teaching and learning techniques, he found a way to drastically improve student performance, helping average students perform better than 98 per cent of their “traditional classroom” contemporaries.
Having discovered a way to improve education efficiency by several orders of magnitude, he could deliver results for students that far outstripped the standards expected of them. However, Bloom also faced a significant challenge – his technique wasn’t scalable.
In 1984 Benjamin Bloom discovered a method for drastically improving educational efficiency, delivering results that are improved by a factor of two standard deviations (two sigma). His method would mean that the “average” student within a given class could now perform better than 49 out of every 50 students within a traditional classroom setting.
Bloom’s secret sauce? The combination of two education approaches: mastery learning and one-to-one tutoring.
- Mastery learning: Each student must achieve true mastery of a topic before moving on to the next, more advanced subject. The student is given the time to study the topic until they succeed in achieving this mastery, even if that takes longer than other students. Time is the one constant in a traditional classroom setting, with lessons squeezed into set periods, leaving no room for different learning rates or individual student needs. With a focus on mastery learning, achievement becomes the constant.
- One-on-one tutoring: Each student is provided with a personal tutor who guides them through their learning, suggesting specific exercises and unlocking the individual student’s potential on an ongoing basis and ensuring they truly “get” the subject.
Given the impact of Bloom’s discovery, one might have expected these techniques to be applied everywhere by now, but this just isn’t the case. Why?
One underlying reason is the resistance to change that can be seen throughout the academic world. Bloom’s new approach to teaching was significantly different to the system in place, meaning that it was naturally met with scepticism and reluctance.
But more important is the time and resource-intensive nature of Bloom’s method. As an approach that’s new to the majority of teachers, substantial time would be needed to set up a mastery-oriented teaching framework, and the move away from hour-long teaching blocks would be hugely disruptive to the traditional learning environment.
Plus, the requirement for one-to-one tutoring would be time and cost-intensive, and incredibly difficult to implement for large groups of students.
To circumvent this, Bloom tried to find other ways to produce his Two Sigma effect, but to no avail. Nothing was as effective as a combination of mastery learning and one-on-one tutoring. The challenge of scalability explains why Bloom’s approach to learning development has not been widely adopted, but that is set to change.
5: What machine learning informs us about human learning?
There are 3 main approaches to training machine learning models: supervised, unsupervised and reinforcement learning. The recent success of AlphaGo zero that learnt to play the game of Go, and defeated the human world champion, without ever being taught how to play the game, shows a lot of promise for self-learning, and developing the skill of learning how to learn. This is also a matter, emphasised in the NEP 2020.
All human learning is — observing something, identifying a pattern, building a theory (model) to explain this pattern and testing this theory to check if its fits in most or all observations.
Overfitting refers to the scenario where a machine learning model can’t generalize or fit well on unseen dataset. Overfitting happens when a model learns the detail and noise in the training dataset to the extent that it negatively impacts the performance of the model on a new dataset.
The opposite of overfitting is underfitting. In human learning overfitting is demonstrated when a learner can do well only on certain types of questions for which s(he) has been very well trained but can’t respond to questions that elicit a broad understanding of the subject.
6: The current state of assessment:
Of the various formats of questions that are used for assessments, multiple choice questions, fill in the blank or match entries from 2 columns are easily capable of being automated and evaluated with computers. Not only can a large number of students be graded automatically, the items can be evaluated on features such as facility, reliability and discrimination index. With good quality items for the bank of questions, it is relatively straightforward to implement an adaptive testing system.
The essay type questions are a formidable challenge to automated machine testing. But recent developments in machine learning, have made grading and assessments of essay type questions possible.
This is also a time that shows up the irrelevance of the monolithic terminal examinations. Sir Anthony Seldon in his 2018 book “ The fourth education revolution” has analysed why such examinations will no longer be relevant. Sir Anthony, Vice-Chancellor of the University of Buckingham, added that continuous assessment would sound the death-knell for exams: “The all-conquering cumulative exam is going to die and we should celebrate its death…The monolithic exam is drawing to a spluttering end.”
Micro-credentials are one of the hot rising idea in the education space. To understand the basics, go look at your child’s Xbox or PlayStation.
For most of the major games, there is an accompanying set of achievements, or badges. Every time a player achieves a particular task (kill 50 zombies without reloading, drive over every tree in the enchanted forest, smash every Lego fire hydrant, etc.) they get a small digital badge on their big page of achievements.
Micro-credentials take a similar approach to education. The root of the idea is simple–you demonstrate a very specific skill, and a badge certifying that micro-credential becomes part of your personal digital file. Some of the earliest micro-credentialing involved computer programming skills, but it has grown far beyond that. To see just how many types of micro-credentials are out there, take a look at Digital Promise.
It offers micro-credentials of its own, but it also provides a platform for other entities to offer their own sets of micro-credentials.
7: AInEd Supremacy: Recently China announced that it had achieved Quantum Supremacy. Google had made the same claim in October2019. The phrase ‘ Quantum Supremacy’ was used in 2012 by the Caltech Professor John Preskill to describe the situation when a Quantum Computer could solve a problem that no classical computer of the day could solve in a reasonable time.
In a similar spirit I have made a phrase “ AInEd Supremacy” to denote the situation when the AI technology of the day in conjunction with a human educator will achieve learning outcomes that no traditional teacher with traditional technology could achieve. We may achieve AInEd Supremacy in about 5 years.
The 2018 book by Anthony Seldon “ The fourth Education revolution” explains how AI will fundamentally transform education. He says that so far education has been the Cinderella of the AI story.
Terrence Sejnowski, Francis Crick Professor at the Salk Institute for Biological Studies and Director of the Computational Neurobiology Laboratory, says that “ Education is going to be the killer app for deep learning”.
Sir Anthony Seldon illustrates how each of five traditional factors in teaching will be transformed by AI over the coming decades:
- Preparation of material will be done by ‘Curation specialists . . whose job it is to work with AI machines to author and identify the most appropriate material for particular student profiles.’
- Organisation of the learning space: ‘Separate classrooms will disappear in time and replaced by pods and wide open, flexible spaces which can be configured for individual and flexible collective learning. Sensors will monitor individual students, measuring their physiological and psychological state, picking up on changes faster and more accurately than any teacher could.’
- Presentation of material to optimise learning/deeper understanding: ‘The flexibility of visual representation with AI allows material to be presented to students which renders much teacher exposition redundant.’
- Setting assignments and assessing/self-assessing progress: ‘Advances in real-time assessment enabled by AI will virtually eliminate this waiting period [the time lag between students being assessed and them receiving feedback on their performance} and ensure feedback comes when most useful for learning.’
- Preparation for terminal examinations and writing summative reports: ‘All this will be swept away by AI. . . . In its place will be attention to continuous data reporting, and real time feedback that will help students discover how to learn autonomously and how to address any deficiencies on their own.’
Will we need teachers in the future? Seldon is clear ‘We do not believe that it is either possible or desirable for AI to eliminate teachers from education’ but he goes on to point out that ‘the application of AI places more responsibility for learning in the hands of the student, for how their time is spent and on what, even from a young age.’ AI will change however the job of the teacher forever. By supporting teaching in all their five traditional tasks, AI will usher in the biggest change the profession has ever seen.’ Interestingly Seldon recognises that remote teaching is a distinct possibility: ‘Imminent advances in virtual technologies will mean too that teachers no longer have to be physically present to offer their services.’
In addition to the help that AI will provide to facilitate the tasks of a teacher, there are a few more results that learning analytics will provide that enhance the teaching quality and the learning effectiveness. For example when the learner is transacting the content, data will automatically be recorded for the time spent at various sections. With permission granted to record facial expressions, there will be very useful data that indicates the level and nature of engagement with the content. The qualities of grit, perseverance etc. referred to by Paul Tough in his seminal work “ How children succeed” can now be observed and analysed quantitatively. Drawing on groundbreaking research in neuroscience, economics, and psychology, Tough shows that the qualities that matter most have less to do with IQ and more to do with character: skills like grit, curiosity, conscientiousness, and optimism.
To help with his class the spring of 2016, a Georgia Tech professor Ashok Goel, hired Jill Watson, a teaching assistant unlike any other in the world. Throughout the semester, she answered questions online for students, relieving the professor’s overworked teaching staff.
But, in fact, Jill Watson was an artificial intelligence bot.
Ashok Goel, a computer science professor, did not reveal Watson’s true identity to students until after they’d turned in their final exams.
With more human-like interaction, Goel expects online learning could become more appealing to students and lead to better educational outcomes.
Just as professors may use AI as teaching assistants, students could use a chatbot to facilitate individual mastery learning, the technology that Bloom suggested in his 1984 research….the two sigma problem. It would have taken about 40 years to realise Bloom’s vision.
In about a decade from now, the impact of the science of learning empowered with Artificial Intelligence will be seen on many features of the present educational landscape that we take for granted. Every student would have access to the best education in the field of choice, in a time schedule that is optimised for the learner. The various School leaving Board examining bodies will become irrelevant, as will the entrance examinations such as the various incarnations of the JEE-IIT, NEET, CAT, LSAT etc. Just like the technologies of CAT-scan, MRI or PET did not replace doctors, but made them more effective, the science of learning with technologies of the 4th Industrial Age will help educators to be a very important element of the success of the country and its people in the emerging knowledge economy.