Forget Making AI Systems That Try To Emulate Humans, How About Human Systems That Try To Emulate AI?

[Warning to readers – this is a bit of a crazy idea. Please forgive me in advance of reading.]

A lot of the narrative about Artificial Intelligence (AI) is about how we can make computer software and robots look and behave like humans. In fact that is one of the two main ways of thinking of AI. (As I mentioned in What I Talk about When I Talk about Artificial Intelligence, I am more interested in the other way, which is just wanting computers to do smart things/ solve difficult problems, not necessarily in the same way humans do.)

But how about this: what can we learn from the way AI works, to change the way humans organize and work. The opposite to what most of the discussion is about. Let me explain my thinking:

The definition of AI is not 100% clear – it’s basically computers doing clever stuff. (Which just defers the definition to ‘what do we mean by clever?’) In the past, expert systems that capture lots of human knowledge in a decision tree, but didn’t learn or adapt, might still be called AI. These days the attention is much more on machine learning (ML), where computers do clever things by learning themselves (as opposed to having all the smarts written in a program by humans).

A particular example of this is Neural Networks (NNs), an approach that was originally inspired by the way with think the human brain might work. A NN is a bunch of very small processing units (neurons) that are all connected to each other. Each neuron does a relatively simple calculation – it takes the output of all the neurons connected to it, giving different weights to each one, then creates its own output based on those inputs. Its output is then, in turn, input to lots of other neurons. The neurons are typically organized into an input layer, one or more hidden layers and an output layer.

For example, the input neurons might be connected to all the pixels in a CCTV camera image, and their might be one output neuron which estimates the likelihood that a crime is about to happen in the shop that the camera is in. The neural network trains itself, typically by using tens of thousands of images that are labelled as ‘crime’ or ‘no-crime’. It does this by changing the weights of the connections between neurons, punishing those that lead to the wrong answer, rewarding the ones that lead to the right answer, over and over again, until they have a network that works well in predicting crimes. The result is a neural network that predicts whether a crime is about to happen based on the image captured by a CCTV camera, hopefully with a decent level of accuracy (e.g. 80-90%).

In the past we might have used an expert system instead of a NN. Analysts would have interviewed store detectives, asked them how they knew where to focus their attention, circumstances that preceded crimes etc. Then they would have encoded that human knowledge into rules and fuzzy logic, for example: “If someone spends more than 5 minutes in any one part of the shop, they are 20% more likely to be about to commit a crime”, or “If someone is wearing a hat that obscures part of their face, they are 10% more likely” etc.

The expert system might perform quite well, but it wouldn’t learn itself, and specifically for the purposes of this article it is not interesting. The expert system represents an encoding of human knowledge, but there is no new knowledge that it creates that we can learn from.

With a neural network, we can learn from its behaviour. For example, the recently well-publicised success of Google Deepmind’s AlphaGo, beating the world champion at the world’s most difficult game, had a couple of examples of AlphaGo (the computer) playing some very surprising moves, that high-level professional Go players wouldn’t have played. After observing them, the humans learned something they didn’t know before. Apparently the two players who had the most contact with AlphaGo went on to play much better themselves. (For more, see: Three Important Things I Learned by Watching the AlphaGo Movie this Week.)

But what I would like to suggest in this article is that there may be something else we can learn from Neural Networks. It is a bit of a complicated idea – please stay with me.

First, let’s imagine that we train a neural network to recognize whether a photograph has an image of a house contained within it. All we do is create a ‘blank’ set of neurons that have inputs from the images, a few hidden layers to process them, then an output that says YES if there is a house in the image, NO otherwise. Then we let it train itself on a large number of images that are labelled YES or NO. Hey presto, within a few minutes or maybe hours, we have this complex software machine that can accurately predict whether an image has a house in it. We haven’t told it anything about how to do that – that’s the bit about neural networks that is thrilling.

Now, let’s imagine that there were 3 hidden layers. When we use tools to examine what each neuron is doing, we will probably find that the first layer of neurons end up taking the picture as input, and identifying basic features. For example, one neuron might have trained itself to find straight edges in the image. Another might have trained itself to find diagonal edges. A third might look for areas that have the colour of bricks etc. So in essence, the first hidden layer of the NN is taking the raw image as input, and then providing a set of features (straight lines, brick-coloured areas etc.) to the second hidden layer of the NN as input. The job of the second layer (which it learned to do itself) might then be to identify areas of the image that look like walls, doors, roofs etc. Finally, the third hidden layer might have trained itself to take walls, doors, roofs etc. as input, and work out if they are in the right configuration to make a house.

This is fascinating stuff, which I am studying in more detail right now, and hopefully excites you about the magic and potential of neural networks too. But here’s the thing I would like to focus on in this article:

(i)  Each neuron becomes a ‘guru’ of some specific task, e.g. identifying straight edges.

(ii) The network becomes a ‘connected set of gurus’ that together solve the problem optimally

(iii)  We never told either the individual neuron what to focus on, or the network how to connect them. They discovered it themselves.

Now let’s use this as an analogy for a human organization (company, government agency, charity, army etc.) Right now, we organize human organization like an expert system, based on a pre-determined structure that we agree in advance. For example – I am going to make a company that sells paint. I am going to structure the company into a bit that buys the chemicals to make paint (Sourcing), a bit that makes the paint (Manufacturing) and a bit that sells the paint (Sales and Marketing). Within the Sales and Marketing division, for example, I am going to have a bit that sells to consumers, a bit that sells to small shops, and a bit that sells to big chain stores. Within the bit that sells to consumers, I am going to have teams that focus on creating advertisements, teams that focus on online eCommerce, and teams that focus on telephone sales. Etc. Etc.

But how about if we tried to emulate neural networks a bit. How about if we made small teams (eg 5-8 people) that we treated as autonomous units in our organization, and tried to create circumstances that allowed them to learn what they could do specially, which other teams (neurons) could help them do that (inputs), and which other teams they could help (outputs).

Obviously, to any sane businessperson, that sounds like an extremely crazy/risky way to run a business, and raises about a million questions, but if we just ‘suspend disbelief’ for a moment, isn’t it an interesting idea? Just like with neural networks, you might get teams focusing on things that you would never imagine separating out into a team, but would be super-powerful.

As a concrete example, I remember years ago meeting the CEO of a European Government’s Collections Agency, whose role was to collect debts unpaid by that country’s citizens. What he told me was that they realized they had become awesome at spotting anomalies in data (e.g. imagine you tell the government you have no money to pay your debt, but your Facebook account shows you went skiing 5 times in a posh resort last year). He told me that this put them in a position that they could spot anomalies in data for other government agencies, and spend more and more of their time and effort on that, and outsource the other aspects of collection (e.g. Printing of documents) to other agencies.

So, we could imagine in future, some very unusual government agencies (e.g. the Ministry of Anomaly Spotting) that when, together with other agencies, formed a much more efficient, effective, learning network of Government.

This is kind of a far-out idea, but if someone thought it ha legs, we could even go beyond using NNs as a metaphor for human organizations. We could actually use NNs to try to simulate an ecosystem, for example a government and its citizens. We could say – let’s imagine we have one agency that is the primary interface to citizens, a hidden layer of agencies that did clever things (like anomaly spotting), and a bunch of output agencies that served the needs of the citizens and the country. We could then set criteria for success, e.g. happiness, health, wealth, then let the network try to experiment and learn, see what it came up with, and try to interpret the results.

What do you think?