The journey from machine learning to true artificial intelligence

In my view we are around half way along the seven level journey from ML to AI and the true potential this can offer to us is still to come. If you listen to the marketeers the world is already full of AI systems, and this is only a half truth. There are now countless machine learning systems out there but real AI systems are a long way off. The difficulties to overcome are real but there are a lot of people working to solve this.

I've been very interested in machine learning and artificial intelligence for nearly 20 years. In fact, my PhD was on this very topic and in a paper written as part of my PhD studies I wrote, back in August 2000, about the potential benefits these will have for humanity - my particular focus was on creating "Next Generation Intelligent Agents" and using these to "help both abled-bodied and disabled people interact with each other across the Internet" where my focus was on how to "advance the field through exploiting natural language processing techniques" with a direction of travel "to develop and integrate the natural language parser part of the software into a new computer operating system and perhaps one day enabling a computer to respond to voice commands making the mouse and keyboard a thing of the past." Well, that world is (finally in my view) very nearly here as covered in last week's post "Voice - its beginning to be everywhere".


This week at the BBC within Platform Engineering I chaired a fantastic discussion on the journey we are on and what comes next after Voice on our journey to true Artificial Intelligence. I've mentioned this journey several times now - about time I explain what mean by it... Basically there are seven fairly distinct levels at which sets of systems can be grouped together each building on the last as we transition from systems that are basic rules based machine learning systems to the potential of hive minds and true artificial intelligence.

1. Rules based systems.

Most “AI” described systems today are really at this level. They are rules based that is they take a set of inputs apply an algorithm to it to determine an outcome. They can get quite complicated but in really are not particularly intelligent. A good example for this is a Recommendations Engine or a simple “Skill” for a voice agent – anything that follows the “if this then that” model.

2. Chat bots and robot processors with context statistic driven.

These are the voice agents of today, web chat agents and arguably basic car systems fall in here. These systems start to have a degree of memory to apply to the rules engines – such as an understanding of specific contexts and remembering the previous outcome as an input to its next decision to make. However, this information is generally transient that is it doesn't "learn" from the information to change future behaviour just to refine the rules at a point in time. A good example of this is “play xyz album”, “volume 3”, “next”, “stop”. As a set of voice input actions that rely on remembering the output from the previous action.

3. Systems with domain specific understanding.

These systems build on the rules sets and contextual awareness to add domain specific knowledge too. Generally they are loaded with a set of understanding on a topic and able to use that detailed knowledgebase to adjust their outcomes. This knowledge base helps the system to narrow down the field of choices and to start to predict forward moves and possible future outcomes as well. Generally built using neural networks and great examples are IBM Watson, DeepMind and Alpha Go.

4. Theory of mind and reasoning machines.

This is the level we hare in practice reached today and where all the research is focused. Its also the dividing line between machine learning systems and true artificial intelligence. This is where systems start to form their own representations of the world instead of being given a set of domain knowledge the systems create their own via trial and error by choosing an outcome and learning (remembering) the result of that outcome – if it was positive or negative and refining their data set accordingly. As a result, they end up needing to “understand” that other inputs can affect the outcomes and also to then make adjustments. To succeed in this area these systems need to be able to model memory – both short term (things I’ve just learned in the current event context) and long term (things I’ve previously learned that I could the system could apply but may not be relevant in the current context) – a very difficult problem to crack; however once done those systems will have their own representations of the world which could be very different to our own.

Only in March this year did DeepMind start to crack thiswith basic robots that "learned" for themselves how to clean-up (move a cylinder to a target area) and receive a reward for it. These are the first machines trying to earn rewards from their own self learned memories! 

And that’s the level we stop at today - right now, there isn’t a robot in existence that could walk/roll into a strange house and using its “learned” knowledge be able to know how to tidy up the environment but, once this problem is cracked the future becomes very interesting.

5. Self aware and artificial general intelligence. (maybe 5 years away)

These systems build on the theory of mind and have a learned awareness of a situation. These are systems that given its learnings determine its own rulesets of success for the future. So instead of being reward or goal driven they make their own choices as to what means success based on both learned and observed outcomes. Ultimately being aware of themselves and that their interactions with an environment also change that environment.

6. Super intelligence. (maybe 10+ years away - close match to young humans)

These are systems that make the change from "I want that" to "I know why I want that". That is these are systems that have a level of consciousness – are aware of themselves, their own existence, have an internal state, and aware of others too and how their interactions with others will have an impact. Those impacts can be both positive or negative and the system makes its own choice as to which way it wants to respond based on its own self learned rule-set.  To do this they need to be able to adapt to their surroundings and the variations along with applying their own learnings and outcomes as predictions on others current or future behaviour (I stop at red lights and I predict others will do too).

7. Singularity - the hive mind of networking. (optimistically 15+ years away - close match to ants/Borg)

This is probably the final level (for now) and this is where these super intelligent systems can share their own knowledge and learnings both passively and actively between each other. As a result, the discoveries of one system change behaviours in others automatically. Ultimately, the network of systems can make collective decisions through combining their knowledge where no one system holds all the information as a result each system within the hive could be removed without the hive collapsing.

In our discussion yesterday we considered how we could utilise systems at various levels within the BBC in the future and how we already do so today. We also had a very interesting discussion about how do we define intelligence and consciousness in the future. There are also the moral issues to overcome, and what role does the BBC have in education and helping to discuss the "scary" factor AI brings about too.

For me personally, given my background its perhaps unsurprising that I love the future prospects we are heading towards. Even though today most systems are in practice only at level 1 or 2, they are capable of some great things - the cost of "moving" up a level isn't yet worth it for the outcomes it might give. As we do move up the levels there are some other challenges to think about too - at the moment rules and regulations haven't even begun to consider all the implications of what happens as we start to reach levels 6 and 7; do these systems start to have a"consciousness" and as such are there are a level of moral implications to consider - can you turn such a system off? Could we build systems to allow the permanent storage of the consciousness? As such could a human mind be stored electronically in the future?