Artificial intelligence (AI) is an emotionally loaded term that strikes fascination into some and fear into others. But if we strip it of fantasy and ignore cyborgs and apocalypse, there is a near-term, practical side of AI that is already unfolding.
Most humans can recognise a chair because they have learnt what a chair is – they can identify thousands of examples of chairs even if they have never seen that chair before. Instead of memorising every image of what a chair could be, humans learn what a chair is and then apply that to new images and examples of chairs.
But how does a computer learn what a chair is? A developer could programme in certain criteria to define the recognition of a chair – shape, size, components or context. But what if the computer had to face new and quirky chair designs? What if it faced a broken chair or a photo taken from a peculiar angle? It would be impossible for a human to programme in every possible variation of a chair and this is where it becomes important for the computer to teach itself – this is machine learning.
The terms machine learning and AI are largely interchangeable. But they are often incorrectly compared with big data and business intelligence (BI). In the example of chair recognition, the computer is given millions of images that contain chairs and millions that don’t. This is big data. Once the software has access to this data, it develops rules as to what makes a chair a chair and not anything else. This is machine learning.
Because the software is working with the pixel-level information of the images and not the social context of what a chair is used for, the rules for recognition developed by the computer could be completely different to those a human would develop.
Leading area
Advanced image recognition is a leading area of machine learning, but it can also be applied to identifying other complex relationships. DataProphet, based in Cape Town, cut its teeth on developing software that optimises the matching of sales agents and possible customer leads in a call centre context – think of the selling of insurance or cellphone contracts via a phone call. Traditionally, potential customer leads would be allocated to agents randomly, but companies quickly discovered that certain agents were better at selling to certain types of customers. The challenge is how best to match them.
Using current BI models, the data would be analysed by comparing previous sales and the characteristics of the agent and customer in each case. As a human writes the algorithms and queries, the parameters need to be limited. For example, the programmer may select age, income level, geographic location, time of day and previous purchase habits. They may find, for example, that younger agents are better at selling to younger customers.
DataProphet decided to apply machine learning models to this challenge and allow the software to use all the information available to develop its own rules for comparison. What it found was that the software broke down the names of agents and customers to discover that similar names led to greater sales conversion. Intuitively, this is because similar names indicate that agents and customers were of a similar culture or language background – information that the company had not initially collected on their customers.
Integrating this information into how customers were matched to agents, DataProphet increased their client’s sales by 34 percent. Machine learning not only learnt the best match on average, but was able to tailor the best match for individual sales agents. The possibilities are endless.
What factors truly cause a change in investor confidence and could machine learning help Pravin Gordhan to streamline his policies and attract foreign direct investment to South Africa? Could machine learning allow banks to go beyond income and assets when determining the risk of default and, therefore, expand lending to the poor at a lower risk? Could machine learning be better than farmers at knowing when to plant, irrigate and harvest? Could it help doctors diagnose disease based on factors that humans have never even thought about?
Machine learning is a blank canvas and human creativity needs to paint it with applications.
* Pierre Heistein is the convener of UCT’s Applied Economics for Smart Decision Making course. Follow him on LinkedIn /in/pierreheistein.
** The views expressed here are not necessarily those of Independent Media.
BUSINESS REPORT