The Leading Edge

The Leading Edge
How state-of-the-art AI systems actually work

If AI is real then the place to find it will be in the world’s most advanced AI systems, which is what we are going to be talking about in the following three podcasts:

IBM Watson: How an oncology instance of IBM Watson was able to detect Melanoma more reliably than the best human dermatologists by Aug-15

Go: How DeepMind’s AlphaGo system bested one of the world’s top-ranked Go players in a three-match tournament  in Mar-16

Dota 2: How a PC-based bot created by OpenAI won a series of games against professional e-gamers who were skilled at playing the video game, Dota 2 in Aug-17

In the previous three Podcasts we looked at three very different AI systems which have been developed for three very different applications.

Taking into account the widespread adoption of AI – in literally 100s of specialist domains – an obvious question arises, which is how might these AI systems be connected together?

For instance, if we think about how the Internet works then we see a large number of nodes (routers) that are connected together in a web and which route packets of data from a source to a destination.

We can imagine replacing the routers in the Internet with domain-specific AI systems and replacing the packets of data with questions and answers.

If we could come up with a way in which such a ‘networked AI’ system might work then we would have seen how to create a single computational structure whose cognitive capability was ‘above human’ in 100s of intellectual domains.

Such a system does not exist, but it might one day. It is worth remembering that telecoms networks, datacommunications networks, mobile phone networks, the Internet and online communities become more valuable with the number of ‘nodes’ that are connected to the network.

So it therefore seems reasonable to assume that the same would apply if we connected together AI systems, which is what we will discuss in the fourth Podcast of this Module:

Networked AI systems: How could we build an artificially intelligent gameplay system that could beat any number of human players at any game?

The three AI systems listed above are unable to do anything other than their intended task, but within their narrow operating domain their apparent cognitive capability already exceeds that of the best human players.

Two of these systems – AlphaGo and OpenAI’s Dota 2 bot – have triumphed at games that the best human players say can only be played well by supplementing years of practice with intuition, instinct and a style of play that does not involve rigorous analysis of move possibilities.

But neither system was specifically engineered to so that it could behave ‘intuitively’.

And perhaps even more surprising is that neither system was programmed with the rules of the games they would play: they ‘learnt’ by playing copies of themselves.

Can we therefore infer that AlphaGo and the Dota 2 bot must possess the machine equivalent of intuition? And, if so, can we infer that these systems must possess a form of real intelligence?

The goal of this module is to answer these questions by opening the lid on these systems and explaining how they work, and why they work.

At the end of the module we will discuss how it might be possible to connect such systems together in a network configuration, which would result in a multi-task, computational structure that could play any game – and beat any human player, or combination of human players.