6 Insights on Artificial Intelligence

At the 44th DNUG conference in Berlin, Tim Bunkus made an excellent introduction to AI and the currently available cognitive capabilities of IBM's Watson service. He inspired me to share my learnings with you. What are your views on AI and cognitive capabilities?

... and thanks to Joerg Raffenbeul and Stefan Gebhart's kind invitation!

Tim Bunkus and Jonas Abel presenting at DNUG '17 in Berlin

1. The big breakthrough: Machines have become very accurate at pattern recognition.

Pattern recognition was for a long time something which humans could do a lot better than machines. 

And then, around 2012, mostly due to the advance of computing power within graphic cards, which enable programmers to make massively parallel linear algebra, we had a number of breakthroughs which now mean that computers have become better than humans at pattern recognition.

This cartoon by xkcd is the most succinct explanation of how cognitive systems works that I've found. 

Behind the terms 'Machine learning', 'cognitive systems' and 'neural networks', it's basically a huge mass of linear algebra that has to be solved in a massively parallel way. Until about 2011, the person on the pile of linear algebra was a human - neural networks were tweaked by humans, with accuracy rates of around 70 - 80%.

But then came a new technology called 'convolutional neural networks', around 2011, which showed incredibly accurate results: error rates lower than 1%. The big change is that the 'person' with the paddle changing the linear algebra is itself a machine. These convolutional neural networks have the ability to self-modify their own parameters in order to improve their accuracy. This is awesome and mind-boggling. 

This is a wonderful video from Computerphile showing the details of this:

2. Humans must train the pattern recogniser. It's time consuming.

It's important to understand the inherent restrictions to the system. Understanding *how* the machines learn, and the fact that the learning process is a tedious, manual one done by humans, should help you see the boundaries of what is possible, and why only a specific subgroup of cognitive tasks can actually be automated.

Imagine the system being just one black box with a very specific type of input, and a single specific output. For instance, take the input to be a picture (any digitised picture of the same size) and the output to be 'true' or 'false', and the skill we want to teach is to recognise if the picture is that of a dog (the example is not chosen arbitrarily), i.e. given any picture, the black box can deliver a 'true' value for 'yes, this is a dog' and 'false' for, 'nope, not a dog'.

The way the machine learns that skill, the 'machine learning' bit, is that we have to feed the machine with a very large quantities of pairs of data, i.e. the input (a picture) and the information 'this is a dog'. In other words, when the machine is learning, we are giving it both its input and the output we expect from it. With a sufficient amount of data, the black box learns, and, astonishingly, can end up being incredibly accurate. In our dog example, this was done in 2012 by the University of Toronto, and they had at their disposal a free database of pictures (Imagenet) with accurate tags as to their content - they could thus automate the machine learning.

That's the nub with these cognitive systems - the assumption is that you need only to give the input, but you actually need to give both an input and an output, and it's the output part which usually implies a lot of manual work.

3. For each input type there is a specific Pattern recogniser.

There are many different flavours of black boxes, each one designed to deal with a particular, quite specific type of input only. One does pictures, another speech recognition, another text ('natural language analyser'). 

The ones that Tim showed IBM Watson Work Service's natural language processing.

These systems are quite large and are for the moment exclusively online, which means that you have to send your data to an external provider.

There is an impressive list of which pattern recognisers are currently being worked on Electronic Frontier Foundations' website: https://www.eff.org/ai/metrics

The one that got me frankly excited is Google Deep Mind's Atari 2600 player. This is a single, universal AI that has taught itself to play any kind of Atari 2600 games, far better than any human:

4. An 'intelligent' system still needs a lot of conventional coding, the 'cognitive magic' is the fantastic pattern recognition

Tim Bunkus showed a particular example where we were training IBM Watson to recognise a request to close a ticket. After a while, the system was recognising the request, i.e. the user wants me to close a particular ticket. What needed to be done after the recognition, i.e. ask for confirmation to the user, actually close the ticket, inform the user that the ticket was used. All this was absolutely standard programming.

This was an eye-opener to me. What was something looking, from the outside, like an intelligent machine that could pass the Turing test, was in fact a hybrid between the natural language pattern recogniser and conventional programming. I imagined it was all really fancy mathematics inside, but Tim made it look pretty straightforward.

5. One Opportunity: We will learn new stuff from machines.

One of the fascinating developments of last year was the perfecting of Google Deep Mind's system to such a point that it could beat Lee Sedol, one of the best Go players in the world. This is a huge deal. Go is a Chinese board game played by successively placing black and white stones on a 19x19 grid. It has been the holy grail of AI for a long time because the range of possible games is several orders of magnitude higher than, say, chess. In other words, it's impossible to brute force the calculations of a Go computer. Even a middling player such as myself could beat Go computers because they were unable to do any kind of pattern recognition. 

Alpha Go run against all the predictions and soundly beat Lee Sedol 4 games out of 5. I followed all games which were being commented by the world's best western Go player, Michael Redmond, and it was fascinating. AlphaGo played mostly a 'quiet' game, not aggressive, and there were times where AlphaGo's moves were unorthodox and surprising. 

At the end of the five games, a somewhat humbled Michael Redmond said that AlphaGo had shown real strength and that human players will have to learn the new, obviously superior playing style of AlphaGo. This has, simply said, enormous implications. Whereas we could before comfort ourselves by saying 'the computer is just giving back what he learned from us', or 'it's just because they are very fast', this is the first instance that I know of that computers have effectively created new knowledge from which we meat-based humans can learn.

Another example of surprising, mind-bending side-effects comes out of Google's translating system. It turns out that the models that the translator uses look like an intermediary, universal language is being created. This is incredible. If we can unlock what that intermediary language is, we could come closer to making accurate language models - a boon for linguists.

This is one of the happiest and most benevolent aspect of artificial intelligence - that we will be able to expand the boundaries of knowledge through artificial intelligence. We're obviously still a long way, because of the very restricted conditions in which we can apply machine learning, but this makes me smile.

6. One Threat: Repetitive cognitive jobs will be soon automated.

Globalisation and automation are the great disruptors nowadays. One of the big surprises is that cognitive jobs who are repetitive and who have a finite set of inputs can be automated too, and the repercussions will be brutal. Professions that were protected for decades, if not centuries, because the skills needed involve a long period of training, are now completely vulnerable to cognitive automation.

The two examples which spring to mind are medicinal diagnosticians and lawyers. In 2007 the Hospital for Sick Children in Ontario started an experiment in their premature birth wards. The at risk babies were connected to about 16 medical devices, continuously monitoring a finite set of measurements. The data collected per day was 1,256 data points collected per second per patient, and this data was just being deleted. The project involved storing the data, which soon reached volumes where big data analysis was possible, and also (and this is the machine learning part of it), with doctors also adding to the data stream any diseases or problems the babies were having. The surprising result: out of all the data collected, the cognitive system, call it 'cogno-doc', was able to accurately identify upcoming diseases up to 48 hours earlier than had been previously possible.

It's a small jump of the mind to imagine that these kind of capabilities, combined with some sort of embedded medical sensors, will be expanded so that the whole job of diagnosing a disease will be done far quicker, earlier, and more accurately by machines instead of doctors.

The same sort of mind experiment could be also applied to jurisprudential lawyers, and I imagine that also such jobs as banking analysts will effectively be replaced by computers.

It's in the nature of software that it can be duplicated very fast. If I create a perfect 'diagno-doc', I can duplicate it thousands of times around the globe (or make the service available via the internet) and the impact will be a sudden, global shock in that industry.

I'm not sure if we should be scared of these developments. In the past, disruptive technological advances were met with anger and even violence, as the demand for a whole profession disappears (I'm thinking of the weavers, the luddites, the saboteurs), but although the transition was tough for those concerned, society as a whole came out the better for it.

On the other hand, there is a potential that this disruption of repetitively cognitive professions will simultaneously affect too large a percentage of the population, i.e. means of production are shifted too fast from persons to computers, and create a sudden burden of unemployment.