Sunday, June 25, 2023

Can machines have empathy and other emotions?

Can machines have empathy and other emotions? Yann Lecun thinks they can and I agree but it is a qualified agreement. This will matter if AI it to become a Universal teacher and have the qualities of an expert teacher.

 

One must start with what emotions are. There has been a good deal of research on this, by Krathwohl, Damasio & Immordino-Yang, Lakoff, Panksepp. Also good work done on uncovering the role of emotion in learning by Nick Shackleton-Jonesl also covered them all in this podcast.


We must also make the distinction between:


Emotional recognition

Display of emotion

Feeling emotions

 

Emotional recognition

The face is a primary indicator of emotions and we look for changes in facial muscles, such as raised eyebrows, narrowed or widened eyes, smiles, frowns, or clenched jaw. Facial scanning can certainly identify emotions using this route. Eye contact is another, a solid gaze showing interest, even anger, while avoiding eye contact can indicate disinterest, shyness, unease or guilt. Microexpressions are also recognisable as expressing emotions. Note that all of this is often a weakness in humans, with a significant difference between men and women, also in those with autism. Emotional recognition is well on its way to being better than most humans and will most likely surpass that ability. 

 

Vocal tone and volume are also significant, tone of voice, intonation, pitch, raised volume when aroused or angry; quiet or softer tone when sad or reflective; upbeat when happy. Body language is another, clearly possible by scanning for folded arms and movements showing unease, disinterest or anger.

 

Even at the level of text, one can use sentiment analysis to spot a range of emotions, as emotions are encoded in laguage. LLMs show this quite dramatically. This can be used to semantically interpret text that reveals a whole range of emotions. It can be used over time, for example, to spot failing students who show negativity in a course. It can be used at an individual level or provide insights into social media monitoring, public opinion, customer feedback, brand perception, and other areas where understanding sentiment is valuable. As it improves, using LLMs it is starting to spot It may struggle with sarcasm, irony and complex language usage.

 

AI already could understands music in some sense, even its emotional intent and effect. Spotify already classify using these criteria using AI. This is not to say it feels emotion.

 

Even at the level of ‘recognition, it could very well be that machine help humans control and modulate bad emotions. I’m sure that feedback loops can calm people down and encourage emotional intelligence. The fact that machines could read stimuli quicker than us and respond quicker, may mean it is better at empathy than we could ever be. Recognising emotion will allow AI to respond appropriately to our needs and should not be dismissed. It can be used as a means to many ends, from education to mental healthcare. Chatbots are already being used to deliver CBT therapy. 

 

Display of emotion

Emotions can be displayed without being felt. Actors can do this, written words in a novel can do this and both can elicit strong human emotions. Coaches do this frequently. Machines can also do this. From the earliest Chatbots, such as ELIZA, that has been clear, Nass &Reeves, showed in 35 studies in The Media Equation, that this reading of human qualities and emotions into machines is common.


As Panksepp repeatedly says we have a tendency to think of emotions as human and therefore ‘good’. Their evolutionary development means they are there for different reasons than we think, which is why they often overwhelm us or have dangerous as well as beneficial consequences. Most crime is driven by emotional impulses such as unpredictable anger, especially violent and sexual crime. This would lead us to conclude that the display of positive emotions should be encouraged, bad ones designed out of the system. There are already efforts to build fairness, kindness, altruism and mercy into systems. It is not just a matter of having a full set of emotions, mort a matter of what emotions we want these systems to display or have.

 

Feeling emotions

This would require AI to be fully embedded in a physical nervous system that can feel in the sense that we feels emotions in the brain. It also seems to require consciousness of the feelings themselves. We could dismiss this as impossible but there are half way houses here and there is another possibility. Geoffry Hinton has posited The Mortal Computer and hybrid computer brain interfaces could very well blur this distinction in a sense of integrating thought with human emotions, in ways as yet not experiences, even subconsciously. But we may not need to go this far.

 

Are emotions necessary in teaching?

I have always been struck by Donald Norman’s argument “Empathy… sounds wonderful but the search for empathy is simply misled.” He argued that this call for empathy in design is wrong-headed and that “the concept is impossible, and even if possible, wrong”. There is no way you can put yourself into the heads of the hundreds, thousands, even tens and hundreds of thousands of learners. As Norman says “It sounds wonderful but the search for empathy is simply misled.” Not only is it not possible to understand individuals in this way, it is just not that useful. It is not empathy but data you need. Who are these people, what do they need to actually do and how can we help them. As people they will be hugely variable but what they need to know and do, in order to achieve a goal, is relatively stable. This has little to do with empathy and a lot to do with understanding and reason.

 

Sure, the emotional side of learning is important and people like Norman, have written and researched the subject extensively. Positive emotions help people learn (Um et al., 2012). Even negative emotions (D’Mello et al., 2014) can help people learn, stimulating attention and motivation, including mild stress (Vogel and Schwabe, 2016). We also know that emotions induce attention (Vuilleumier, 2005) and motivation that can be described as curiosity, where the novel or surprising can stimulate active interest (Oudeyer et al., 2016). In short, emotional events are remembered longer, more clearly and accurately than neutral events.

 

All too often we latch on to a noun in the learning world without thinking much about what it actually means, what experts in the field say about it and bandy it about as though it were a certain truth. But trying to induce emotion in the teaching and design process may not be not that relevant or pnly relevant to the degree that mimicing emotion may be enough. AI can be designed to induce and manipulate the learner towards positive emotions and not the emotions, identified by Panksepp and others, that harm learning, such as fear, anxiety and anger. We are in such a rush to include ‘emotion’ in design that we confuse emotion in learning process with emotion in the teacher and designer. It also seems like lazy signalling, for not doing the hard analysis up front, defaulting to the loose language of concern and sympathy.

 

Conclusion

In discussion emotions we tend to think of it as a uniquely human phenomenon. It is not. Animals clearly have emotions. This is not a case of human exceptionalism. In other words, beings with less complexity than us can feel. At what point therefore can the bottom up process create machine that can feel? We seem to be getting there and have come quite far having reached ‘recognition’ and ‘display; 

 

If developments in AI have taught us one thing, it is to never say never. Exponential advances are now being made and this will continue, with some of the largest companies with huge investments, along with a significant shift in research and government intentions. We already have the recognition and display of emotions. The feeling of emotions may be far off, unnecessary for many tasks, even teaching and learning.

 


In medicine, empathy is already being helped with GPT4, patients can benefit from being helped by both a knowledgeable and empathetic machine. We see this already Healthcare in the Ayers (2023) research, where 79% of the time, patients rated the chatbot significantly higher for both quality and empathy. That’s before the obvious benefits of being available 24/7, getting quicker results, increased availability of healthcare in rural areas, access by the poor and decreased workload for healthcare systems. It empowers the patient. For more on this area of AI helping patients with empathy listen to Peter Lee’s excellent podcast here. He shows that even pseudo-empathy can run deep and be used in many interaction with teachers, doctors, in retail and so on.

This is why I think the Universal Teacher and Universal Doctor are now on the horizon.

 

Bibliography

Ayers et al. 2023. Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Platform

Norman, D.A., 2004. Emotional design: Why we love (or hate) everyday things. Basic Civitas Books.

Norman, D., 2019. Why I Don't Believe in Empathic Design.

Um, E., Plass, J.L., Hayward, E.O. and Homer, B.D., 2012. Emotional design in multimedia learning. Journal of educational psychology104(2), p.485.

D’Mello, S., Lehman, B., Pekrun, R. and Graesser, A., 2014. Confusion can be beneficial for learning. Learning and Instruction29, pp.153-170.

Vogel, S. and Schwabe, L., 2016. Learning and memory under stress: implications for the classroom. npj Science of Learning1(1), pp.1-10.

Vuilleumier, P., 2005. How brains beware: neural mechanisms of emotional attention. Trends in cognitive sciences9(12), pp.585-594.

Oudeyer, P.Y., Gottlieb, J. and Lopes, M., 2016. Intrinsic motivation, curiosity, and learning: Theory and applications in educational technologies. Progress in brain research229, pp.257-284.

https://greatmindsonlearning.libsyn.com/affective-learning-with-donald-clark 

Friday, June 23, 2023

Is the new Digital Divide in AI between the 'EU' and 'Rest of the World'?

OpenAI has opened its first foreign office in London citing the pro-innovation economy, talent and, it is clear although not stated, the fear of EU regulation. They are also clearly cautious about EU regulation. Bard was available in 180 countries and territories, including the UK, but NOT the EU, until a deal was done. Facebook has been holding back releases of models, Twitter has left the EU’s voluntary code of practice. Is this a new Digital Divide? One wonders what effect this will have on investment in AI across the EU? The lack of debate around the consequences of this is puzzling. 

When Italy declared UDI and banned ChatGPT they quickly relented (actually a move by a right wing appointee to show their strength). But this is different, large AI providers, such as Google, and Facebook, are taking the initiative and simply not releasing AI services in EU countries. This new Digital Divide may soon be between the EU and the rest of the world and could have serious consequences.

 

On the other hand, the EU is a huge and wealthy market, so the large tech companies will not take these decisions lightly. The problem is that the EU's legislation is often bureaucratic and cumbersome, involving lots of paperwork, hits on productivity. and is a stick and not carrot mechanism. The famous pop-up consent solution ‘Manage all cookies’ is GDPR nonsense, as no one reads the consent forms, it is therefore largely a waste of time. It was the result of bad legislation, producing a massive hit on productivity with no tangible benefits. One side overlegislates, the other is perhaps too lax and defensive – the net result is a Digital Divide.


With the release of Baidu's Ernie 3.5, which is neck to neck with ChatGPT4 on performance, this has turned into a two horse race - US and China. There's a third horse, but that's a moral high horse, which has barely left the stalls - that's the EU.

 

Economic environment

Let’s start with the big picture. 


In 2008 the EU economy was larger than America’s. In 2008 the EU’s economy was nearly 10% larger than America’s at $16.2tn versus $14.7tn. By 2022, the US economy had grown to $25tn, whereas the EU and the UK together had only reached $19.8tn… Now the US is nearly one-third bigger. It is more than 50 per cent larger than the EU without the UK…” (FT 20 June 2023) and that gap is growing. 


The US has trounced Europe in terms of productivity, economic growth, investment models, research, investment, the creation of tech companies, defence and energy policy. It has also trounced the EU in terms of AI research and implementation. If the EU cannot develop a strong tech-based economy it will have to rely on low growth legacy markets, such as tourism and luxury goods, meaning it will fall further behind.

 

Productivity deficit

AI matters as ‘productivity’ needs a well-educated and skilled workforce, good infrastructure, and a favourable business and investment environment. Importantly, those with the more sophisticated tools tend to be the more productive. The evidence for the productivity gains using AI, within just a few months, is clear. If the EU either ban such tools or create an environment where angels fear to tread, then productivity on coding, management and general output, as these tools affect almost every sector, will start to lag. We had a dry run when Italy banned ChatGPT and there were reports of falls in productivity.

 

Training and education deficit

The University research and teaching system that feeds AI tech in terms of core research and skilled labour is dominated by the US, UK and China. EU Universities barely figure in the major rankings. An additional problem is the now deeply rooted anti-corporate sentiment in Higher Education in the EU. The sneering attitude towards the private sector, even OpenAI as a not-for profit, is now the norm, often accompanied by a failure to understand its actual structure. A symptom of this is that the debate in Higher Education focused largely, not on learning, but plagiarism. Far too little debate has taken place on the benefits in education and health.

 

Effort in AI is skewed towards often vague ethical initiatives making the overall atmosphere one of negativity, slowing down progress. The danger is that the benefits will be realised elsewhere while the EU remains rooted in old, analogue instititutions, where everything in tech is seen as a moral problem. There is nothing wrong with the moral debate but it is so often driven by fearmongering and activism, and not objective moral debate, which is to look at the moral issues and consequences, good and bad, not just the bad.

 

An additional problem is the unlikely adoption of AI in education in the EU. The real initiatives such as Khan Academy and Duolingo have been funded and implemented in the US, aided by philanthropic investment. There is little of that energy and type of investment in Europe. As AI becomes integrated into education and training in the US, its absence here will mean less productivity. 

 

Investment freeze

Investors looked askance at Italy’s surprise ban and widened their astonished gaze across the whole of the EU. If one country can do this, so can others. Investors have a currency – it is called ‘risk’. They assess and quantify risks and base decisions based on that risk analysis. Anyone who has been through the process knows that they do their homework and due diligence. One of those risks is already baked in, the Italy ChatGPT ban, huge punishment fines is another, the generally negative rhetoric and cultural context is yet another. Why would large scale investors pump cash into a territory where bans, fines and an absence of services have become the norm? Investors like a favourable business environment not one that is based on negativity and punishment.

 

Investment model

The model that emerged post-war in the US has proved superior to that in the EU – private sector, investors, government and Universities working together on large projects with a real focus on impact. In AI we now see the fruits of that system in the US, where ground-breaking research on foundation models take place in large tech companies and not-for-profits, such as OpenAI. Europe scoffs, and gets bogged down in long-winded, bureaucratic and low impact Horizon projects, while the US and China gets on with getting things done. In truth we now look to the US for investment and that is the market most want to expand in, as it is the largest growth market on the planet. They have become so dominant that they merely buy European companies in AI.

 

Ethical quicksand

Generative AI has launched a thousand quangos, groups and bad PhDs on ‘AI

and ethics’ across Europe. You can’t move for reports, frameworks and ideas for regulations which rain down on us from publicly funded organisations, with far too little attention on potential solutions and benefits. 

 

Debate on the benefits has been swept aside by pontificating and grand-standing. It is easy to stand on the sidelines as part of the jeering, pessimist mob, less easy to do something positive to actually solve these issues. Rather than solve the problems of safety, security, alignment and guard-railing with real solutions, the EU has chosen to see the glass, not as half full, but as brimming with hemlock. It sees laws and fines as the solution, not design and engineering.

 

The EU also has no laws banning VPNs and their use is becoming more common. This is a huge loophole when using AI services. It is already happening with Bard, as the internet is like water, it tends to seep round and into places, based on demand.

 

Punishment strategy

The EU has been issuing fines for some time now, although not is all as it appears. You may note that many of these large fines are issued from Ireland, but there has been a long and bitter fight between Dublin and Brussels. Ireland gains a good portion of its GDP from a small number of US tech companies and because they are based there, the GDPR fines come from there. They have fought these fines tooth and nail but, in the end, had to bend the knee to Brussels.


 

There is something reasonable in these latest fines as the data transferred may be used by US surveillance agencies (they have a bad track record here). In practice Meta have until later in the year to comply. This is a bit of a cat and mouse game, with politics at the heart of it all. On the other had the EU puts up with Ireland and Luxembourg stealing other countries tax revenues through massive tax evasion. It is all a bit of a tangled mess.

 

The bottom line, however, is that this tactic is resulting in a deeper rift. Twitter quit the EU’s voluntary code of practice in mid-May as it could be fined up to 6% of its global revenue (£145m) or be banned across the EU if it does not comply with the Digital Services Act. Facebook are making noises about abandoning the EU.

 

Solutions

If as much effort went into solutions, than regulations, fines and rhetoric, we would progress at the right pace, solving problems as we go, rather than trying to punish people into submission. Hacker-led safety testing, well funded research, effort on international ISO standards not regional efforts with a focus only on large operation and parameter models and implementations would all help. Above all third-party professional hacker teams can be deployed to identify security and data weaknesses before release. Incident reporting can also be useful. This collaborative, non-confrontational approach is far preferable to the negativity and sledgehammer of legislation and punishing fines.

 

Conclusion

These battles have been raging for some time, mostly behind the scenes but Google and Facebook have also had run ins with Canada and Australia. Some of this has been resolved, some not. There is something predictable about it all - the old world versus the new.


It is an inconvenient truth but the EU is too late to the party, the US and China have forged ahead in IT and AI with their own tech giants. The EU has failed to create tech giants and has also deliberately chosen the path of being some sort of global regulator but it has a weak economy, weak research, weak investment and a weak entrepreneurial culture. In the same way that the Ukraine war showed the EUs lack of investment in defence and a lack of any overall defence policy, where not for the first time it had to rely on the US to come to its aid and provide arms, cash and expertise, so it is with AI. 

 

The investment is low, there is no policy other than taking a morally superior stance. So much energy has gone into ethical hand-wringing that Europe is reduced to being a bystander. It thinks it has sway but it is a fraction of the world’s population, shrinking, and white Eurocentrism now seems more than a little dated. It rides its lumbering moral high-horse, looking down on the rest of the world, while others like the US feed it a little hay to keep it happy and speed past. It would surely be better developing AI solutions that have identified benefits in productivity, learning and healthcare, than simply regulating it.

Thursday, June 22, 2023

Pask, Conversational Theory & Generative AI

An often forgotten learning theorist, who needs to be revisited in the light of Generative, conversational AI, is Gordon Pask (1928 – 96). His work on learning machines and conversational theory, although reaching back to the 1950s, has turned out to be both prophetic and useful. 

Known as the ‘Dandy of Cybernetics’ he was famously mannered, theatrical, intense and eccentric. With his cape, Edwardian suit, bow-tie and pipe, he was known as being difficult to communicate with, his lectures and writing often difficult to comprehend. Both an academic and entrepreneur (largely unsuccessful) he liked the freedom to build and experiment and valued his autonomy. 

 

Nevertheless, he was an original thinker who has made a significant contribution to thinking about the complexity of learning. A polymath interested in geology, engineering, art, theatre, sculpture, biological computing, artificial intelligence, cognitive science, logic, linguistics, psychology music, cybernetics and education, he saw learning as a complex process of interaction.

 

Well aware of Artificial Intelligence, he was a friend of Marvin Minsky and would stay with him in the US, although he was critical of the AI community’s tendency to focus on isolated systems, separate from human interaction. As AI has widened out into language models and become part of the global system of learning, Pask has come back into focus, with his sophisticated theories of learning through conversation.

 

Learning machines

Like Pressey and Skinner before him, he developed a number of learning machines. The difference was in seeing machines as things we converse with, rather than use as tools. In this sense he was closer to Pressey than Skinner. Eucrates, for example, not only taught but could adapt its teaching based on knowledge of the learner. There is physical feedback, and the machine builds a model of user behavior. He drew upon Wittgenstein’s idea of meaning as use. This is in stark contrast to Skinner’s behaviourist reward machines.

 

Developed in the 1953s, he invented Musicolour, a system of interaction between performer and system that allowed a light show to be triggered by a musician. It could amplify a performance, even become ‘bored’ by the musician’s repetition! It went on tour around Music Halls in England and was interaction as performance.

 

A more practical learning machine was SAKI (Self-Adaptive Keyboard Instructor), built in 1956, to teach punch-card operators, with electric wires that cued the learner, detected keypresses and time intervals, always pushing the learner ahead but not too far ahead, to keep the learner motivated. It focused on the intense and fruitful interaction between learner and machine. It was still in use in the UK Post Office in the late 1960s.

 

Ecogame, developed in 1969, a management training game, had some success being bought by IBM. Targeted at business and political leadership, it was based on his ideas of self-organized, interactive learning in a simulated process and environment.

 

His teleprinter trainer caste CASTE, which stood for Course Assembly System and Tutorial Environment, was the size of a small room, built in 1972. It was an early adaptive learning system based on his learning theory. CASTE had an Entailment Structure or topic map to provide an overview, a Communications Module on a computer terminal, BOSS ( Belief and Opinion Sampling System) and a modelling system that would present learning and assessment tasks. Systems were designed for students with tutorial algorithms, to spot failure and reassess on lower levels of learning. It offered a range of choices to the learner. 

 

Thoughtsticker, built around the same time, used entailment meshes, was a hierarchical representation of knowledge. It originally had a set of pigeon holes with paper notes, later developed into an adaptive system that allowed users to move upwards on an entailment mesh of concepts. You could study it through a mesh of concepts with different perspectives. The learner would feed in concepts and the interconnections would be shown. A small computer was later used. It could also jump with was conditional hyperlinking. It was ahead of its time, unfortunately it was also mired in learning styles theory.


In all of this development and experimentation he was trying to implement his theory of learning - conversation theory. Teaching and learning were fundamentally conversations fo Pask, even when they involved technology.

 

Conversation theory

Pask studied learning through conversation, looking for identifiable features processes. Meaning is agreed use in the context of a conversation. This is not a fixed exchange of propositions but an exchange that shapes the thinking and learning of both parties. We may agree, disagree, modify beliefs, even change our minds but we gain from the conversation. Learners and teachers move through knowledge changing perspectives and levels. This, for Pask was the essence of learning. He sees conversation and learning as something active, in use, as we deal with the world. Conversation is also the link or knowledge we have with others, that they exist, like us. Conversation for Pask is at the root of everything we do, perception, emotional engagement and language. This is a much subtler idea than just verbal conversation. 


We learn together and are always using artefacts to learn and communicate. A conversation can be several different types, as an individual, two people, a group of people or combination of people and technology. Proximity is not the point, as technology destroys distance. This is a general conversational theory that sees people, their social interactions and technology as one.


Conversational theory saw learning as taking place in a learning environments, where we had sophisticated encounters with that environment, to it and back from it. We learn, Pask thought, by interacting, or having ‘conversations’, with people and our environment, including gestures, pictorial, with media and with and through machines. In that sense ‘conversation’ is almost metaphorical. We interact with the world in a way that is very much like ‘conversation’ in language.


At one level, teacher and learner can engage in a strict conversation, with interactions on two levels – how (teacher demonstrates) and learner does and why (explains) at the conceptual level. Pask saw that we engage in hierarchies of conversation but not in some simplistic Bloom hierarchy, we flit between levels. What he called styles of teaching or learning are like Wittgenstein’s language games, contextually different forms of conversation.

 

His learning theory saw conversation as the essential learning process. He studied conversations separately from the content of conversations to explore their processes and limits. Learning conversations are, and should be, open, complex and dynamic, not a clumsy, didactic movement toward a final truth. Variety, a plurality of ideas, perspectives and conversations matters, where learning emerges from the process of conversation.

 

Learning needs to provoked by a teacher and constructed by the learner and although Pask is rarely seen as a serious social constructivist, he did a lot of work, unlike most constructivists, on exactly how he though learners constructed knowledge and pragmatic skills. Importantly, he was sensitive to learning by doing and practical skills as well as knowledge.

 

In Conversation, cognition and learning (1975) he outlined the complexity of conversation, especially in learning, with hierarchical branching and loops, agreement, understanding, along with analogies and generalisations. Concepts are refined, expanded, altered, generalised, applied, related by analogy in the process of learning. This is not just dry rationalism, as for Pask to think is to feel, mind and body intertwined. He was keenly aware of our evolutionary legacy as biological beings. Conversation is how we think, and to teach and learn we must reproduce the process that evolution has bequeathed us. He had drawn from Vygotsky, the idea of a social being having developed throughout childhood, although Pask's focus was less on language, more on a wider social context.

 

Entailment meshes

Entailment meshes attempt to capture thought processes, with topics related to each other and grouped into relations, such as ‘analogies’ and ‘coherences’, that comprise concepts, which we share. We are the active constructers of concepts, making new finer distinctions, seeing similarities and differences, generalised, realised through action and interaction. All of this takes place naturally, we are conversational beings and have no choice in the matter. Conversation is an on-going, dynamic process, part of our being.

 

This is a radical departure from the database driven model that drove AI for so long, one of retrieval. We now interact with the world to learn through conversation. These conversations can be varied as they must suit the progress of the learner, not be pre-determined by too few constricting rules. He thought that most learning technology was hopelessly primitive in teaching and learning as they failed to recognise conversation as the means by which we learn. 


In learning there also has to be a contract to learn, common ground and action for successful learning, a process that is honest about contradictions, conflict, epiphanies and resolutions in the learning process. It is an agreed context.


Once a domain has been mapped out, one can also record the learner’s vectors through content, the routes taken, nodes reached. This is the sort of personalisation realised by modern adaptive systems and now generative AI, that can deliver the sort of sophisticated dialogue that Pask envisaged.

 

Interactions of actors

Pask and Gerard de Zeeuw took conversation theory one step further with the ‘Interactions of actors’ theory, where the scope was widened to include three or more actors in conversation, across time, entering and leaving such conversations. Here the defined process had knots, links, braids and a logic of the conversational and learning process but this is often metaphorical and lacks the complexity that AI models now have in capturing language. Rather then designing such a logic of process, AI learns from language and its use in Large Language Models, where the training data set is enormous. It is therefore empirically much more useful and representative of the reality of language use.

 

Generative AI

Conversations go on over time and become the culture, so LLMs capture many conversations. Conversations are captured as culture. So rather than constructing huge logic trees and cumbersome Paskian machines, Generative AI seems to deliver the richness of topics and language that allows it to deliver at least some of the conversational complexity that Pask thought was essential for learning. It already holds the relationships of words and phrases to each other by being trained in an unsupervised way, with a vast corpus of text, then further supervised training using RLHF (Reinforcement Learning from Human Feedback) and PPO (Proximal Policy Optimisation). Pask even talked about evil as the limits of the rights of actors to interact or hold conversation, prefiguring alignment and guard-railing.

 

The next stage is to provide some conversational structures to the dialogue that allow efficient learning to take place. This is precisely what is happening with Khanmigo, where structured feedback and encouragement adds motivational and useful feedback to the on-going learning conversation. Pask would have been delighted at the emergence of this type of learning system using Generative AI. 

 

Interaction is still primitive with computers, and we may have just seen a more conversational model emerge, where conversation, let’s call it on-going prompting, is replacing search and retrieval or request systems. Systems are becoming conversational, whether reading/writing or oral/aural and therefore more human. The technology must become more aware of people's needsto have more fruitful conversations. Conversations encapsulate interests, the seeking and learning process. They reveal intention.

 

The word ‘conversation’ may seem odd in the case of LLMs, as it lies beyond the idea of discrete individuals and stored knowledge into massive and sophisticated models with billions of parameters trained on unimaginably large sets of training data. But our interactions with these systems are certainly conversational, as learning is a process not an event. Pask realised that despite the radically different way these models work, they are still entities within a conversation with humans. The biological and technical are not separate but entwined in a single conversational system.

 

Conclusion

We would do well to pay more attention to learning through Pask’s conversational theory, with more empirical work on what works well in the context of increasing capability on the machine side through AI. Improvements that take the learning process into real conversations with structures that both allow for complexity but also guide towards outcomes, is now badly needed. This is what will give us useful, massively beneficial, teaching and learning. Let’s call them conversational systems, even a Universal Teacher.


Bibliography

Pask, G. (1961). [1968]. An approach to cybernetics. Hutchinson. 

Pask, G. (1962). a proposed evolutionary model. in H. von Foerster & G. W. J. Zopf (Eds.), Principles of self-organization: Transactions of the University of Illinois symposium on self-organization, Robert Allerton Park, 8 and 9 June 1960 (pp. 229–254). Pergamon Press. 

Pask, G. (1966). Men, machines and the control of learning. Educational Technology, 6(22), 1–12. 

Pask, G. (1968). a cybernetic model for some types of learning and mentation. in H. l. oestreicher & d. r. Moore (Eds.), Cybernetic problems in bionics (pp. 531–586). Gordon and Breach science Publishers. 

Pask, G. (1975). Conversation, cognition and learning: A cybernetic theory and methodology. Elsevier. 

Pask, G. (1975). The cybernetics of human learning and performance: A guide to theory and research. Hutchinson Educational. 

Pask, G. (1976). Conversation theory: Applications in education and epistemology. Elsevier.

Pask, G., & Curran, s. (1982). Microman: Living and growing with computers. Century Publising Co. 

Pask, G., & Kopstein, F. F. (1977). teaching machines revisited in the light of conversation theory. Educational Technology, 17(10), 38–41. 

Pask, G., & von Foerster, H. (1960). A predictive model for self-organizing systems.

High-horses are all very well but not if you want to get somewhere

Bard is available in 180 countries and territories, including the UK, but not the EU. Is this a new Digital Divide? One wonders what effect this will have on investment in AI across the EU? Other AI services are also not releasing. The lack of debate around the consequences of this is astounding. 

When Italy declared UDI and banned ChatGPT they quickly relented (actually a move by right wingers to show their strength). But this is different, large AI providers, such as Google and Facebook, are taking the initiative and simply not releasing AI services in EU countries. This new Digital Divide may soon be between the EU and the rest of the world and could have serious consequences.

 

On the other hand, the EU is a huge and wealthy market, so the large tech companies will not take these decisions lightly. The problem is that its legislation is often bureaucratic and cumbersome, involving lots of paperwork and hits on productivity. The famous pop-up consent solution ‘Manage all cookies’ GDPR nonsense (no one reads the consent forms, therefore largely a waste of time) )was the result of bad legislation, producing a massive hit on productivity with no tangible benefits.

 

Economic environment

Let’s start with the big picture. “In 2008 the EU economy was somewhat larger than America’s. In 2008 the EU’s economy was nearly 10% larger than America’s at $16.2tn versus $14.7tn, by 2022, the US economy had grown to $25tn, whereas the EU and the UK together had only reached $19.8tn… Now the US is nearly one-third bigger. It is more than 50 per cent larger than the EU without the UK…” (FT 20 June 2023) and that gap is growing. The US has trounced Europe in terms of productivity, economic growth, investment models, research, investment, the creation of tech companies, defence and energy policy. It has also trounced the EU in terms of AI research and implementation.

 

Productivity deficit

All of the above matters as ‘productivity’ needs a well-educated and skilled workforce, good infrastructure, and a favourable business and investment environment. Importantly, those with the more sophisticated tools tend to be the more productive. If the EU either ban such tools or create an environment where angels fear to tread, then productivity on coding, management and general output, as these tools affect almost every sector, will fall behind. We had a dry run when Italy banned ChatGPT there were reports of falls in productivity.

 

Training and education deficit

The University research and teaching system that feeds AI tech in terms of research and skilled labour is dominated by the US, UK and China. EU Universities barely figure in the rankings. An additional problem is the now deeply rooted anti-corporate sentiment in Higher Education in both the UK and EU. The sneering attitude towards the private sector, even OpenAI as a not-for profit, is now the norm.

 

Effort in AI is skewed towards ethics making the atmosphere negative and progress slow. This is a great danger that the benefits will be realised elsewhere while we remain rooted in our Medieval institituions seeing everything as a moral problem. There is noithing wrong with the moral debate but it is so often driven by fearmongering and acticism, with little actual moral or ethical foundation, which is to look at the moral issues and consequences, good and bad.

 

An additional problem is the unlikely adoption of AI in education. The real initiatives such as Khan Academy and Duolingo have been funded and implemented in the US, aided by philanthropic investment. There is little of that energy in Europe. As AI become integrated into education and training in the US, its absence here will mean less productivity in learning. 

 

Investment freeze

Investors looked askance at Italy’s surprise ban, and widened their gaze across the whole of the EU. If one country can do this, so can others. Investors have a currency – it is called ‘risk’. They assess and quantify risks. One of those risks is already the Italy ban, huge fines are another, the general rhetoric and cultural context is yet another. Why would large scale investments flow into a territory where bans, fines and an absence of services have become the norm? Investors like a favourable business environment not one that is based on punishment.

 

Investment model

The model that emerged post-war in the US has proved superior to that in Europe – private sector, investors, government  and Universities working together on large projects with a real focus on impact. In AI we now see the fruits of that system in the US, where ground-breaking research on foundation models takes place in large tech companies and not-for-profits, such as OpenAI. Europe scoffs, and gets bogged down in long-winded, bureaucratic, vague and low impact Horizon projects, while the US gets on with getting things done. In truth we look to the US for investment and that is the market most want to expand in as it is the largest growth market on the planet. They have become so dominant that they merely buy European companies in AI.

 

Ethical quicksand

Generative AI has launched a thousand quangos, groups and bad PhDs on ‘AI

and ethics’ across Europe. You can’t move for reports, frameworks and regulations which  rain down on us from publicly funded organisations, with far too little attention on potential solutions and benefits. The debate in Higher Education focused largely. Not on learning, but plagiarism. Far too little debate has taken place on the benefits in education and health,

 

Debate on the benefits has been swept aside by pontificating and grand-standing. It is easy to stand on the sidelines as part of the jeering, pessimist mob, less easy to do something positive to actually solve these issues. Rather than solve the problems of alignment, guard-railing with real solutions, the EU has chosen to see the glass, not as half full, but as brimming with hemlock. It sees laws and fines as the solution, not design and engineering.

 

The EU has no laws banning VPNs and their use is becoming more common. This is a huge loophole when using AI services. It is already happening with Bard as the internet is like water, it tends to seep round and into places, based on demand.

 

Conclusion

It is an inconvenient truth but the EU is too late to the party, the US and China have forged ahead in IT and AI with their own tech giants. The EU has deliberately choosing the path of being some sort of global regulator but it has a weak economy, weak research, weak investment and a weak entrepreneurial culture. In the same way that the Ukraine war showed the EUs lack of investment in defence and a lack of any overall defence policy, where not for the first time it had to rely on the US to come to its aid and provide arms, cash and expertise, so it is with AI. 

 

The investment is low, there is no policy other than taking a morally superior stance. So much energy has gone into ethical hand-wringing that Europe is reduced to being a bystander. It thinks it has sway but it is a fraction of the world’s population and shrinking, and white Eurocentrism now seems more than a little dated. It rides its lumbering moral high-horse, looking down on the rest of the world, while others like the US feed it a little hay to keep it happy and speed past. It would surely be better developing AI solutions that have identified benefits in productivity, learning and healthcare, than simply regulating it.

 

 

Monday, June 19, 2023

Personalised tutors - a dumb rich kid is more likely to graduate from college than a smart poor one

A dumb rich kid is more likely to graduate from college than a smart poor one and traditional teaching has not solved the problem of gaps in attainment. Scotland, my own country, is a great example, where the whole curriculum was up-ended and nothing has been gained. In truth, we now have a solution that has been around for some time. I have been involved in such systems fo decades.


AI chatbot tutors, such as Khanmigo, are now being tested in schools but not for the first time. The Gates Foundation has been at this for over 8 years and I was involved in trials from 2015 onwards, and even earlier with SCHOLAR, where we showed a grade increase among users.. We know this works. From Bloom’s famous paper onwards, the simple fact that detailed feedback to get learners through problems they encounter as they learn works. 


“It will enable every student in the United States, and eventually on the planet, to effectively have a world-class personal tutor” says Salman Khan. Gates, who has provide $10 million to Khan agrees, “The AIs will get to that ability, to be as good a tutor as any human ever could” at a recent conference.

 

We see in ChatGPT, Bard and other systems increased capability in accuracy, provenance and feedback, along with guardrailing. To criticise such systems for early errors now seems churlish. They’re getting better very fast.

 

Variety of tutor types

We are already seeing a variety of teacher-type systems emerge, as I outlined in my book AI for Learning.


Adaptive, personalised learning means adapting the online experience to the individual’s needs as they learn, in the way a personal tutor would intervene. The aim is to provide, what many teachers provide, a learning experience that is tailored to the needs of you as an individual learner.

The Curious Case of Benjamin Bloom

Benjamin Bloom is best know for his taxonomy of learning (now shown to be weak and simplistic), wrote a far less read paper, The 2 Sigma Problem, which compared the lecture, formative feedback lecture and one-to-one tuition. It is a landmark in adaptive learning. Taking the ‘straight lecture’ as the mean, he found an 84% increase in mastery above the mean for a ‘formative feedback’ approach to teaching and an astonishing 98% increase in mastery for ‘one-to-one tuition’. Google’s Peter Norvig famously said that if you only have to read one paper to support online learning, this is it. In other words, the increase in efficacy for tailored  one-to-one, because of the increase in on-task learning, is huge. This paper deserves to be read by anyone looking at improving the efficacy of learning as it shows hugely significant improvements by simply altering the way teachers interact with learners. Online learning has to date mostly delivered fairly linear and non-adaptive experiences, whether it’s through self-paced structured learning, scenario-based learning, simulations or informal learning. But we are now in the position of having technology, especially AI, that can deliver what Bloom called ‘one-to-one learning’.

Adaption can be many things but at the heart of the process is a decision to present something to the learner based on what the system knows about the learners, learning or context.

 

Pre-course adaptive

Macro-decisions

You can adapt a learning journey at the macro level, recommending skills, courses, even careers based on your individual needs.

 

Pre-test

‘Pre-test’ the learner, to create a prior profile, before staring the course, then present relevant content. The adaptive software makes a decision based on data specific to that individual. You may start with personal data, such as educational background, competence in previous courses and so on. This is a highly deterministic approach that has limited personalisation and learning benefits but may prevent many from taking unnecessary courses.

 

Test-out

Allow learners to ‘test-out’ at points in the course to save them time on progression. This short-circuits unnecessary work but has limited benefits in terms of varied learning for individuals.

 

Preference (be careful)

One can ask or test the learner for their learning style or media preference. Unfortunately, research has shown that false constructs such as learning styles, which do not exist, make no difference on learning outcomes. Personality type is another, although one must be careful with poorly validated outputs from the likes of Myers-Briggs, which are ill-advised. The OCEAN model is much better validated. One can also use learner opinions, although this is also fraught with danger. Learners are often quite mistaken, not only about what they have learnt but also optimal strategies for learning. So, it is possible to use all sorts of personal data to determine how and what someone should be taught but one has to be very, very careful.

 

Within-course adaptive

Micro-adaptive courses adjust frequently during a course to determine different routes based on their preferences, what the learner has done or based on specially designed algorithms. A lot of the early adaptive software within courses uses re-sequencing, this is much more sophisticated with Generative AI. The idea is that most learning goes wrong when things are presented that are either too easy, too hard or not relevant for the learner at that moment. One can us the idea of desirable difficulty here to determine a learning experience that is challenging enough to keep the learner driving forward.

 

Algorithm-based

It is worth introducing AI at this point, as it is having a profound effect on all areas of human endeavour. It is inevitable, in my view, that this will also happen in the learning game. Adaptive learning is how the large tech companies deliver to your timeline on Facebook/Twitter, sell to you on Amazon, get you to watch stuff on Netflix. They use an array of techniques based on data they gather, statistics, data mining and AI techniques to improve the delivery of their service to you as an individual. Evidence that AI and adaptive techniques will work in learning, especially in adaption, is there on every device on almost every service we use online. Education is just a bit of a slow learner.

 

Decisions may be based simply on what the system thinks your level of capability is at that moment, based on formative assessment and other factors. The regular testing of learners, not only improves retention, it gathers useful data about what the system knows about the learner. Failure is not a problem here. Indeed, evidence suggests that making mistakes may be critical to good learning strategies.

 

Decisions within a course use an algorithm with complex data needs. This provides a much more powerful method for dynamic decision making. At this more fine-grained level, every screen can be regarded as a fresh adaption at that specific point in the course.

 

AI techniques can, of course, be used in systems that learn and improve as they go. Such systems are often trained using data at the start and then use data as they go to improve the system. The more learners use the system, the better it becomes.

 

Confidence adaption

Another measure, common in adaptive systems, is the measurement of confidence. You may be asked a question then also asked how confident you are of your answer.

 

Learning theory 

Good learning theory can also be baked into the algorithms, such as retrieval, interleaving and spaced practice. Care can be taken over cognitive load and even personalised performance support provided adapting to an individual’s availability and schedule. Duolingo is sensitive to these needs and provides spaced-practice, aware of the fact that you may have not done anything recently and forgotten stuff. Embodying good learning theory and practice may be what is needed to introduce often counterintuitive methods into teaching, that are resisted by human teachers. This is at the heart of systems being developed using Generative AI, the baking-in of good learning theory, such as good design, deliberate practice, spaced practice, interleaving, seeing learning as a process not an event.

 

Across courses adaptive

Aggregated data

Aggregated data from a learner’ performance on a previous or previous courses can be used. As can aggregated data of all students who have taken the course. One has to be careful here, as one cohort may have started at a different level of competence than another cohort. There may also be differences on other skills, such as reading comprehension, background knowledge, English as a second language and so on.

 

Adaptive across curricula

Adaptive software can be applied within a course, across a set of courses but also across an entire curriculum. The idea is that personalisation becomes more targeted, the more you use the system and that competences identified earlier may help determine later sequencing.

 

Post-course adaptive

Adaptive assessment systems

There’s also adaptive assessment, where test items are presented, based on your performance on previous questions. They often start with a mean test item then select harder or easier items as the learner progresses. This can be built into Generative AI assessment.

 

Memory retention systems

Some adaptive systems focus on memory retrieval, retention and recall. They present content, often in a spaced-practice pattern and repeat, remediate and retest to increase retention. These can be powerful systems for the consolidation of learning and can be produced using Generative AI.

 

Performance support adaption

Moving beyond courses to performance support, delivering learning when you need it, is another form of adaptive delivery that can be sensitive to your individual needs as well as context. These have been delivered within the workflow, often embedded in social communications systems, sometimes as chatbots. Such systems are being developed as we speak.

 

Conclusion

There are many forms of adaptive learning, in terms of the points of intervention, basis of adaption, technology and purpose. If you want to experience one that is accessible and free, try Duolingo.


ASU trials

Earlier trials, with more rules-based but sophisticated systems proved the case years ago. AI in general, and adaptive learning systems in particular, will have enormous long-term effect on teaching, learner attainment and student drop-out. This was confirmed by the results from courses run at Arizona State University from  2015. 

One course, Biology 100, delivered as blended learning, was examined in detail. The students did the adaptive work then brought that knowledge to class, where group work and teaching took place – a flipped classroom model. This data was presented at the Educause Learning Initiative in San Antonio in February and is impressive.

Aims
The aim of this technology enhanced teaching system was to:
increase attainment
reduce in dropout rates
maintain student motivation
increase teacher effectiveness


It is not easy to juggle all three at the same time but ASU want these undergraduate courses to be a success on all three fronts, as they are seen as the foundation for sustainable progress by students as they move through a full degree course.

1. Higher attainment

A dumb rich kid is more likely to graduate from college than a smart poor one. So, these increases in attainment are therefore hugely significant, especially for students from low income backgrounds, in high enrolment courses. Many interventions in education show razor thin improvements. These are significant, not just on overall attainment rates but, just as importantly, the way this squeezes dropout rates. It’s a double dividend.


2. Lower dropout 

A key indicator is the immediate impact on drop-out. It can be catastrophic for the students and, as funding follows students, also the institution. Between 41-45% of those who enrol in US colleges drop out. Given the 1.3 trillion student debt problem and the fact that these students dropout, but still carry the burden of that debt, this is a catastrophic level of failure. In the UK it is 16%. As we can see increase overall attainment and you squeeze dropout and failure. Too many teachers and institutions are coasting with predictable dropout and failure rates. This can change. The fall in drop out rate for the most experienced instructor was also greater than for other instructors. In fact the fall was dramatic.


3. Experienced instructor effect

An interesting effect emerged from the data. Both attainment and lower dropout were better with the most experienced instructor. Most instructors take two years until their class grades rise to a stable level. In this trial the most experienced instructor achieved greater attainment rises (13%), as well as the greatest fall in dropout rates (18%).

4. Usability

Adaptive learning systems do not follow the usual linear path. This often makes the adaptive interface look different and navigation difficult. The danger is that students don't know what to do next or feel lost. In this case ASU saw good student acceptance across the board. 



5. Creating content
One of the difficulties in adaptive, AI-driven systems, is the creation of ustable content. By content, I mean content, structures, assessment items and so on. We created a suite of tools that allow instructors to create a network of content, working back from objectives. Automatic help with layout and conversion of content is also used. Once done, this creates a complex network of learning content that students vector through, each student taking a different path, depending on their on-going performance. The system is like a satnav, always trying to get students to their destination, even when they go off course.

6. Teacher dashboards

Beyond these results lie something even more promising. The  system slews off detailed and useful data on every student, as well as analyses of that data. Different dashboards give unprecedented insights, in real-time, of student performance. This allows the instructor to help those in need. The promise here, is of continuous improvement, badly needed in education. We could be looking at an approach that not only improves the performance of teachers but also of the system itself, the consequence being on-going improvement in attainment, dropout and motivation in students.

7. Automatic course improvement
Adaptive systems take an AI approach, where the system uses its own data to automatically readjust the course to make it better. Poor content, badly designed questions and so on, are identified by the system itself and automatically adjusted. So, as the courses get better, as they will, the student results are likely to get better.

8. Useful across the curriculum
By way of contrast, ASU is also running a US History course, very different from Biology. Similar results are being reported. The platform is content agnostic and has been designed to run any course. Evidence has already emerged that this approach works in both STEM and humanities courses.

9. Personalisation works
Underlying this approach is the idea that all learners are different and that one-size-fits-all, largely linear courses, delivered largely by lectures, do not deliver to this need. It is precisely this dimension, the real-time adjustment of the learning to the needs of the individual that produce the reults, as well as the increase in the teacher’s ability to know and adjust their teaching to the class and individual student needs through real-time data.

10. Student’s want more

Over 80% of students on this first experience of an adaptive course, said they wanted to use this approach in other modules and courses. This is heartening, as without their acceptance, it is difficult to see this approach working well.



Conclusion

    We have been here before. These systems work. They are already powerful teachers and will become Universal Teachers in any subject. The sooner we invest and get on with the task, the better for learners.