Embrace Complexity (Part 4)
Embedding intelligence into your organisation with network-shaped AI
Data, Cloud and AI are three fundamental forces that are fuelling an exciting (and at times scary) technological phase transition where we all move from the smog of the industrial age into the neon glow of the information age. If you have been following this series then you will know that we have been using the network shape like a north star that guides us as we search for a unified theory that connects these three forces together. A unified theory that can be used by all organisations so that fewer of them are left behind.
The graph-shaped data introduced in part 2 provides us with networked information and networked information captures a richness and complexity that can only be found in the relationships that connect the parts together. The nuance and subtlety that exists in the lines connecting the dots.
The graph-shaped cloud of computers introduced in part 3 adds in networked distribution and that allows ALL data to be connected no matter where it is stored. This gives ‘normal’ organisations access to a vast sea of connected information.
Now we move on to the final and perhaps most exciting piece of the puzzle as graph-shaped AI adds networked intelligence that can move beyond learning in ‘traditional’ euclidean straight lines allowing computers to begin thinking outside of the box.
Current AI
AI is a broad subject but for our purposes here we are interested in one branch, namely artificial neural networks. An artificial neural network mimics the brain through layers of interconnected nodes known as artificial neurons. Each node will fire based upon the strength of the connections flowing into it. Artificial neural networks learn by running over input data many times adjusting the strength of the connections until the model learns the ‘weightings’ that will predict the right answer.
Deep neural networks, stack these networks into many layers and are able to provide state-of-the-art performance in an increasing number of fields. They are actively being used to recognise our faces and speech with almost scary accuracy. But the key question to ask here is: why can’t most existing organizations harness the power of these deep neural networks?
Why is it so hard for ‘normal’ organisations to get their AI projects to do anything that actually impacts their bottom line?
Networked AI
The architectural north star of networkification provides a two-part answer to this question. Firstly, most organisations don’t have enough volume of data and the data service addresses this issue by connecting thousands of existing fragmented internal data sources together into a single vast and distributed graph hosted in the cloud.
Secondly, there is also a more fundamental problem. Although neural networks are graph-shaped, we have been training them on rectangular, box-shaped data. This means that the neural networks inevitably learn this linearity and return rigid box-shaped answers back to us.
Take another look at the current neural network architecture. Can you spot those tell-tale horizontal straight lines between the layers? Once again it is a case of box-shaped linearity hiding in plain sight right before our eyes.
This rectangular linearity means that most of the current AI models are learning about the values of the data but they have no explicit information about the connective structure of the systems in which those values exist. The models have no context and therefore they have no sense of the bigger picture.
Next time you talk to a computer on the phone and grow frustrated at its lack of flexibility, its inability to step out of its rigid tracks, reflect on the straight lines highlighted in the image above.
So let's apply systems thinking to AI. Artificial neural networks are obviously network-shaped and with the help of a graph adapter so is our data. Therefore we can perform our second unification and merge the predictive capability of AI directly into the wider structure of the data-cloud network itself.
Each node in the network is now three things at once. It is a data point, and it is a network address in the cloud, and it is a node in an AI model. Each node contains the three forces of data, cloud and AI united as one.
With this unification in place, we can send nuanced, curvy, network-shaped data into our neural networks and get sophisticated, context-sensitive, network-shaped answers back out!
Graph Convolutional Networks
How can this be done? Well, there is a new breed of graph-based neural networks that work natively with network-shaped data. They are called Graph Convolutional Networks (or GCNs for short) and they learn not just about the parts but also the connections between them as well. GCNs do not seem to need vast datasets but can make quite intuitive leaps based upon smaller sets of connected information.
GCNs use a process called message passing where each node makes predictions and then passes those predictions to its nearest neighbours in the network. This process is then repeated so that each node can take the predictions of its neighbours into account in the next iteration.
(As a side note, rather than using the node’s nearest neighbours I personally favour what I call a ‘pathway’ based version of message passing but that is a whole other article series).
In short we can say that GCNs use message passing to learn a model that is able to make predictions that are based upon the connective structure of the whole system and this is how each node in the graph gets a sense of the bigger picture.
Applications that you use each day, such as the route finding on Google Maps and the restaurant recommendations on Uber Eats, are now powered by these graph-based machine learning algorithms. All of the tech giants are investing heavily in this area and many of us believe GCNs will come to dominate the near future of AI.
But GCNs are also of particular relevance to ‘normal’ organisations whose available training data is not made up of huge sets of the same thing (like millions of pictures of cats scraped from the web) but rather is made up of a diverse and distributed graph of many interconnected internal applications and databases. Therefore, daring organisations can choose to skip out much of the earlier AI (which they seem to be able to make little practical use of anyway) and concentrate on GCNs.
With this unification, each data point is a node in the graph and can have a mirrored AI twin that is a node in the model. So each node can be trained to make intelligent predictions about novel situations. This enables AI to be incorporated into all levels of an organisation’s operation and delivers systemic intelligence and automation.
Furthermore, the weightings that inform this intelligence are learned not just from the values of the nodes in the ‘data network’, but also from the connective shape of the surrounding network itself.
You can think of these weightings as being like a set of coordinates that point to the location of a given node on a graph-shaped conceptual map. The weightings allow a computer to calculate the ‘conceptual distance’ between two nodes.
The graph adapter gives us rich network-shaped data, the data service gives us a vast sea of connected data on the cloud and now the GCNs can learn insights directly from this rich, distributed network and make systemic predictions about the future. This is computers thinking in curves and it is hard to convey how exciting the potential of this really is.
Embedding Your Knowledge
Furthermore, the GCN weightings (known as node embeddings) can be stored back into the graph as pure data.
The embeddings from the GCNs model are stored back as part of the data cloud by saving it as another three-part statement in the graph. In this way, each node can begin to build up a set of embeddings that reflect the various models that it has been a part of. This is the other side of the unification as the embeddings are stored back directly into the unified data network. The last piece of the puzzle has been slotted into place and our long journey searching for a unified theory is coming to an end.
Teaching AI How to Generalize
Of particular importance to the intelligence of the unified network are the nodes in the shared conceptual model. These abstract concepts will appear in many different models and can therefore collect lots of node embeddings from a variety of diverse situations. My early research shows that this seems to be able to teach an organisation’s AI a basic form of generalization. In this sense, it is possible that the conceptual model could eventually grow into something like the organisation’s brain.
Bringing Your Organisation to Life
Networks have an organic structure that appears throughout the natural world, your brain is a network of nerve cells, your circulatory system is a branching network of capillaries, trees and plants are also branching networks and their roots are then woven together in vast lacy networks of mycelium.
It appears networks are life’s go-to structure for processing information and you could even say that life itself is network shaped.
It should therefore come as no surprise that we can make our AI more lifelike by using the network shape to unify data, cloud and AI into one organic structure. In essence, networks can bring our organisations to life.
All the hard work that would have to be put in by an organisation's staff:
- standing up data services to share datasets on the cloud
- using graph adapters to create nuanced graph-shaped data structures
- connecting the datasets together with unique network addresses
- and conforming to a shared conceptual model
All this hard work to structure the organisation's collective knowledge representation now really pays off because the graph-shaped AI can walk along the many pathways that criss-cross the vast organisation-spanning knowledge graph and provide systemic insights.
The GCN is the active ingredient that breathes life into the network and allows AI to make context-sensitive predictions that impact an organisation’s bottom line. The GCN is the final practical tool that is needed to realise a unified theory of network-shaped technological phase transition.
Tool Number Three: GCNs
Each data service should mirror the passive graph-shaped data with an active graph-shaped machine learning model that gives each node the potential to also learn and predict. The data service should publish these learned node embeddings back into the network as pure data.