What we talk about when we talk about AI
About a month ago Lee Sedol, formerly the second-ranking Go player in the world, retired from professional play. Citing the existence of “an entity that cannot be defeated” as the reason for his departure, Lee’s announcement comes three years after being defeated 1-4 in a series of matches played against a machine: Deepmind’s AlphaGo. The version of AlphaGo that beat Lee Sedol in 2016 used a machine learning algorithm (specifically an artificial neural network) trained using reinforcement learning techniques on a large dataset of human games to “learn” winning strategies. AlphaGo’s successor, AlphaGo Zero, was given nothing but the rules of Go. After playing itself over three days, it beat the original AlphaGo 100-0.
With what may be a healthy dose of hyperbole, Demis Hassabis, co-founder and CEO of Deepmind, claimed that AlphaGo Zero’s power lay in the fact that it was “no longer constrained by the limits of human knowledge”. In its second game against Lee Sedol, AlphaGo played a move that many found bewildering in its sheer elegance and creativity. Expert Go player Fan Hui remarked, “It’s not a human move. I’ve never seen a human play this move.” Strikingly, he also described the move as “so beautiful”.
When we talk about artificial intelligence (AI) in 2019, whether it be self-driving cars, voice assistants, cancer-diagnosing machines or the genius humbling AlphaGo, we are invariably talking about the family of mathematical models called machine learning. Ash Priest, managing partner at Novigi, recently had the pleasure of chairing an innovation discussion group on behalf of the Association of Superannuation Funds Australia (ASFA). In it we discussed the current capability of machine learning, and the possible application of its increasingly superhuman powers in superannuation.
Machine learning is an umbrella term that includes a very broad range of mathematical models, of wildly varying complexity. Perhaps the simplest algorithm that is technically (and commonly referred to as) machine learning is simple linear regression. If you have ever added a trendline to a scatter plot in Excel, feel free to go ahead and add “Machine Learning” to your list of skills on LinkedIn. On the other end of the scale we have deep learning algorithms: artificial neural networks (ANNs) in their many varieties, which are loosely based (read: very loosely) on the structure and function of the human brain.
Machine learning algorithms are suited to tasks in which:
- They can be taught by example — without necessarily requiring explanation.
They can be left unsupervised to find patterns in data — which would often not be obvious to a human.
They can work within a defined set of rules and parameters to achieve a desired outcome.
The internet is brimming with in-depth explanations of how various machine learning algorithms actually work, so we’ll refrain from recycling this information here. We will however, take a brief look at a vanilla neural network to try and highlight some key considerations in determining the applications of machine learning algorithms more broadly.
Example: Vanilla Neural Network for Image Classification
Let’s consider a relatively simple example of a neural network, that we want to train to determine whether an image contains a picture of a dog, a cat, or neither. The network is composed of a series of interconnected neurons — represented here as coloured dots.
The input layer takes in the image data, with each neuron being activated with a value between 0 and 1 (corresponding to the brightness of each pixel in black and white image, for example). In essence, each neuron in subsequent layers takes in one or more inputs, performs a mathematical function on that input, and outputs it to neurons in the next layer (or as a result, if the neuron is in the final layer).
In the diagram above, which could represent any of the neuron in the 2nd hidden layer, we see that the inputs interact with a weight (denoted as w1, w2, w3, w4). Usually this is multiplicative, i.e. 0.7 x w1, with the products of each input and its associated weight then summed together with a bias parameter (denoted as b1) which controls how sensitive the individual neuron is to previous inputs (with an adjustment to bring the output of the function between 0 and 1).
When we talk about training the neural network, what we’re actually doing is finding values for these weights and bias parameters such that the model classifies images in our training set with the highest possible accuracy. This is generally done by using techniques like gradient descent and backpropagation to find local minima of a loss function (for those interested in the maths, please see source). These techniques are designed to train neural networks efficiently, which is important because the number of parameters grows quickly as the number of neurons and layers in the network increases. As an example, the simple image classification network described above would have a total of (50 x 50 x 4) + (4 x 4) + (4 x 3) = 10,028 weight parameters and 4 + 4 + 3 = 11 bias parameters if it was built and trained to detect images with 50×50 pixel dimensions.
Though other machine learning algorithms differ greatly from the example above, there are some common characteristics worth noting. All machine learning algorithms have some concept of a loss function, and are “trained” by minimising this function. All machine learning algorithms operate under the assumption that there are underlying relationships in your data to be uncovered. In situations where data is completely uncorrelated, a sufficiently complex model can still be fit to a training set. As the model presupposes that there are inherent relationships in that data, if there turn out not to be, it will have no predictive power.
With some basic understanding of what machine learning is and what it can be used for, we turn to superannuation. While in previous innovation discussion groups we sometimes struggled to find applications for some overhyped technology, with machine learning and artificial intelligence, it is clear that there are a huge number of possibilities. Obviously not every avenue explored will lead to a successful implementation, but we are confident that many will. Here we outline just a small fraction of the many ideas floated in the last discussion group.
Marketing and Member Engagement
An increasing trend — typically in industries other than superannuation — is to take a data-driven (machine learning) approach to the challenge of attracting and satisfying customers. The marketing and engagement process can be thought of as a cycle where customers are firstly segmented into appropriate sub-groups based on shared characteristics. Relevant engagement and advertising strategies can then be developed for each of these segments. Finally, the effectiveness of these strategies can be measured by looking at the impact to key outcomes. The results of the final step can then be used to make adjustments and incrementally optimise the process going forward.
It is critical that superannuation funds think carefully about how to approach each of these steps and take learnings from other industries so that they can accurately tailor strategies to their members that will deliver on many of the desired outcomes like increase member engagement and satisfaction.
A brief description of how machine learning techniques can be applied to each of these steps is covered below.
Member segmentation – Rather than making assumptions around how the current and potential future membership base could be segmented based on traditional variables such as age, gender and income, a fund could instead use unsupervised machine learning techniques like clustering to automatically determine segments from the data. This reduces bias and could lead to interesting and unintuitive member segments being generated that end up consisting of individuals that respond in similar ways to different strategies. There also exists a range of scoring methodologies that can be used to measure the accuracy of the segmentation process which can help with the selection of an appropriate algorithm.
Marketing attribution – Marketing attribution is the process of identifying the set of actions from an individual that contributed to a desired outcome. In the context of superannuation, this could be determining the relevant contribution of different advertising sources in encouraging a member or employer to join the fund. Once these are known (or approximated), the fund can then optimise its marketing mix to generate the greatest number of desirable outcomes (e.g. new members) per dollar spent.
The process of attribution is not simple, however, because not all member actions are tracked, and even if they are it isn’t clear what the relative contribution of each action should be. This is where Machine Learning can help. Machine Learning models can look at all action and outcome data (both internal and external to the fund) and use this to more accurately determine these weightings.
Member engagement – It is important for a fund to determine how engaged a member is because accurate measurement is the first step to improvement. If a measure of engagement can be determined for each member (i.e. a Member Engagement Score (MES) ∈ [0 ,1]) then a fund can use this to do things like predict and reduce churn by devoting marketing effort towards members who have a low MES.
There are rudimentary ways of determining this engagement score by looking at a series of potential indicators like time spent on a website or actions like investment switches. Instead, more sophisticated supervised machine learning techniques can be used. A Machine Learning model can be trained using known historical data such as members who have left the fund as they are likely to have low engagement before leaving. This model can then be used to more accurately determine an engagement score for a member or flag members that are at a high risk of churning. The machine learning model can use a wide range of data known about a member for prediction including sentiment inferred from unstructured data sources.
The same approach can also be used to measure other things about a member like satisfaction or loyalty.
Risk Management and Compliance
Risk management and compliance is a perpetual concern for super funds, and an area in which the application of machine learning and artificial intelligence shows particular promise. There are two key scenarios in which we think AI and ML could add a great deal of value, reflecting current issues in the superannuation system:
Identifying human error – In both our personal experience with funds and in our conversations with risk and compliance experts in superannuation, human error was identified as a key source of risk. Data entry and processing issues in — for example — the processing of contributions or the payment of benefits are common. These issues usually go unnoticed until recognised by members, if they are recognised at all.
Identifying unusual or fraudulent behaviour – Superannuation is the sixth most popular target for fraudsters and con-artists in Australia. There are at any given time a number of scams in operation, that result in members being conned out of their benefits.
A supervised machine learning algorithm has the potential to be of assistance here. Errors and discrepancies in the data, relating to either human error or fraud, could be flagged and added to a continuously growing training set. This training set could then be used to train a machine learning solution, that would then be able to identify data points that have a high probability of being the product of either human error or fraudulent behaviour. Importantly, it could feasibly do this in real time and stop new data issues from being introduced. Armed with this information, a human operator could then identify the source of the discrepancy and take steps to remediate.
Artificial intelligence has established itself as a force for disruption in the modern economy. The sheer volume of AI applications that have already become commonplace in the lives of millions of people are a good indication that there’s more to it than simply hype. Superannuation is not special, and there will be a competitive edge for those funds and service providers who manage to integrate AI into their business models. AI and machine learning — and data science more broadly — are complex and intimidating fields, and the superannuation sector will need to access expertise it does not currently have. Finding good partners will be key.
Terry Donnelly leads the Data Services function at Novigi, and is based in the Sydney office.
Kevin Fernandez leads the consulting business at Novigi, and is based in the Melbourne office.
For more information about anything you’ve read here, or if you have a more general inquiry, please contact us.