Saturday, October 08, 2016

Collaborative AI is learning from shared experience – collective intelligence is here and it’s terrifying

AI is moving fast, scoring one victory after another in specific domains. However, the main problem AI has is in moving from one domain to another. It may be great at playing chess or GO, or other rules based games, but when it comes to other simple but different problems, it is not flexible. This problem, getting AI to be more general in terms of its skills or learn to apply what it learned in one domain to another, is a serious limitation, perhaps the greatest limitation of current AI.
We humans have a different but no less debilitating problem – our brains. We literally have to spend up to 20 years or more in classrooms being painstakingly taught by other humans to acquire knowledge and skills. Even then, it’s only a start on the long road that is lifelong learning. That’s because we cannot efficiently transfer knowledge and skills directly from one brain to another. It cannot be uploaded and downloaded. In addition to this limitation, we forget most of what we are taught, sleep 8 hours a day, are largely inattentive for much of the remaining 16 hours, get ill and die. Artificial intelligence has none of these constraints.
Cloud robotics
One solution to the learning problem in AI, now being practised in robotics, is ‘cloud robotics’, where one robot can literally ‘teach’ another. By teach, I mean, pass and share its acquired skills on to other robots. This is a bit mind blowing, as it is something we cannot do as humans.
Google and others have been experimenting with cloud robotics for some years, where robots learn how to do something, through neural networks and reinforcement learning (trial and error) and once they have acquired that skill, it can be uploaded to as many other robots as you want. They literally share experiences and therefore learning. Not only do the sets of robot learn quickly, they instantly share that learning with other networked robots. This whole idea of AI learning from collective or shared experience is fascinating.
Google have been research three different (but not mutually exclusive) techniques for teaching or training multiple robots collectively, by allowing them to share experiences and learn general purpose skills:
1. learning motion skills directly from experience
2. learning internal models of physics
3. learning skills with human assistance
Shared experience, in all three of these forms, clearly takes less time than a single robot acquiring its own experiences. But it’s not only time that shrinks, you also get the benefit of variation in those experiences, more diversity of experience. Deeper learning, in terms of both quantity and quality, takes place that can cope with more complex environments and problems. This sharing of experience, whether trial and error tasks, models that are built or human data that is used to train robots can be networked, shared and seen as a form of pooled, collective intelligence.
Why is this frightening? Robots are already decimating the manufacturing sector. The possibility that generalpurpose robots will be more flexible and able to do all sorts of tasks as they learn from each other, is frightening in the sense that this takes them beynd the specifics of manufacturing.
Collective intelligence
Collective intelligence, a term coined by Pierre Levy, was defined by him as,
a form of universally distributed intelligence, constantly enhanced, coordinated in real time, and resulting in the effective mobilization of skills"
But Levy’s theory of collective intelligence is now dated and inadequate. Firstly, it has an inadequate definition of ‘collective’ that has been superseded by recent developments, not only in social media but in other forms of technology such as AI. Secondly, it has an inadequate definition of ‘intelligence’ that has been superseded by recent thinking about ‘networks’ and ‘intelligence’.
Networks and intelligence
More attention needs to be given to the nature and role of ‘networks’ in collective intelligence. Some philosophers posit the idea that networks are intelligent, to a degree, simply by virtue of being networks. Our brains are networks, indeed the most complex networks we know of, and artificial intelligence uses that same (or similar) networked power to interact with our brains. We do not learn in a linear fashion, like video recordings nor do we remember things alphabetically or hierarchically. Our brains are networks with pre-existing knowledge, and intelligence, that needs to fit with other forms of knowledge from networks.
The theory of connectivism, proposed by Stephen Downes and George Siemens posits ‘Connectivism’, as a theory where “knowledge is distributed across a network of connections, and therefore that learning consists of the ability to construct and traverse those networks”. It is an alternative to behaviourism, cognitivism and constructionism. ‘Connectivism’ focuses on the connections, not the meanings or structures connected across networks. Intelligence, existing and acquired, is the practices, by both teachers and learners, that result in the formation and use of effective networks with properties such as diversity, autonomy, openness, and connectivity. This challenges the existing paradigms, that do not take into account the explosion of network technology, as well as presenting a new perspective on collective use and intelligence. Connectivism can also be used to bring in newer technological advances and newer agents - such as artificial intelligence.
Koch argues that the line for consciousness and intelligence has changed to include animals, even insects, indeed anything with a network of neurons. He takes includes any communicating network. We have evidence that consciousness and intelligence is related to networked activity in both organic brains and non-organic neural networks. Could it be that intelligence is simply a function of this networking and that all networked entities are, to some degree, intelligent?
AI and collective intelligence
There are several recent technological developments that open up the possibility of collective intelligence. The most important new technologies is AI (Artificial Intelligence). Artificial Intelligence is embedded in many online media experiences. This takes us beyond the current flat, largely text-based, hyperlinked world of text and images, such as Wikipedia, or even social media, into forms of media that are closer to Lévy’s original idea of collective intelligence. Networks store knowledge but, with the advent of online AI, they can also be said to BE intelligent. AI is intelligence that resides in a network and is intelligent ‘in itself’ but also adds to the sum of collective intelligence when used by humans.
Machine learning (code that learns and creates code), with the aid of collective human and machine created data, actually becomes more ‘intelligent’. The more it is used the more intelligent it becomes, sometimes surpassing the intelligence of humans, in specific domains. As we saw at the start of this article, it can now also be shared. This is a new species of shared intelligence that is shared in realtime, direct  and scalable.
Collective Artificial Intelligence, let’s call it CAI, can be said to reside in and be an emergent feature of all networks, human and artificial, organic and non-organic. In other words, the agents of collective intelligence have to be widened, as does the nature of that intelligence and the interaction between them all.
We are only just beginning to see and practically explore and build new forms of collective intelligence that allow teaching and learning to be done by machines very quickly and on a massive scale. This is exhilarating and frightening at the same time. We, as humans, are now part of a networked nexus of human teachers, human learners, AI teachers, AI learners, networked knowledge, and networked skills. The world has suddenly become a lot more complex.

(Thanks To Callum Clark for the ideas on Levi and collective intelligence.)

No comments: