Listen to this article
Deep learning may be one of the most overhyped of modern technologies, but there is a good chance that it will one day become the secret sauce in many different business processes. For anyone entering the workforce now — or thinking about how to position their career for the long term — this would be a very good time to understand its implications better.
The term “deep learning” refers to the use of artificial neural networks to carry out a form of advanced pattern recognition. Algorithms are trained on large amounts of data, then applied to fresh data that is to be analysed. It has become the hottest subject in the field of artificial intelligence, thanks in particular to breakthroughs in image and language recognition in recent years that have approached or surpassed human levels of comprehension.
The potential scale of deep learning’s impact on business was laid out last month in a report from McKinsey Global Institute, Notes from the AI Frontier: Insight from Hundreds of Use Cases. Depending on the industry it is in, the value a company could hope to gain from applying this technology ranges from 1 to 9 per cent of its revenues, according to the consultants.
This points to trillions of dollars of potential impact on business — and the workers who are the first to learn how to apply it will be the big winners, according to Michael Chui, a McKinsey partner.
“If you learn sooner and faster, you have the chance to do much better relative to others,” he says. That applies not only to people with technical skills, but to any manager who works out how to use the technology to tackle business problems.
The reason that managers who learn how to make use of the technique have the chance to leap ahead of others, Chui explains, is that technologies such as deep learning can have an outsized impact throughout a business. “The technologies are levers of value creation . . . digital means you can do more, faster. If you can successfully scale something across an organisation or a customer base, you have that much more impact.”
The best way to think of deep learning is as a form of advanced analytics. Given enough data to train the algorithm, it can be used in many different tasks. The challenges include identifying the types of problem that are most susceptible to being solved with this technique, picking the particular approach that is best in any given situation and making sure the algorithms are fed with a good supply of high-quality and timely data.
Most of the business potential in deep learning, according to McKinsey, will come in two broad areas: marketing and sales, and supply chains and manufacturing. Examples of the former include customer service management, creating individualised offers, acquiring customers and honing prices and promotions.
This means that companies in consumer industries stand to benefit more than most from deep learning. Frequent interactions with customers generate the kind of data needed to feed the systems. Using real-time data to predict demand trends on a hyper-regional basis can add 0.25-0.75 per cent to sales, the consultants estimate, with further benefits from lower waste and spoilage.
Uses of the technology in supply chains and manufacturing, meanwhile, include predictive maintenance of equipment, yield optimisation, procurement analytics and inventory optimisation.
These potential benefits are purely theoretical, based on the capabilities of the technology, and it will take some time for most companies to capture them. There is, for instance, a huge skills shortage in the business world. Data scientists and machine-learning specialists are now the most in-demand IT experts, based on the high salaries they are attracting.
The tools that non-expert developers need to make use of the technology are also still in their infancy. Many of the recent advances in the field still count as cutting-edge research. Services such as Google’s Cloud AutoML represent the first real attempt to make deep learning more widely available as a practical tool, by automating parts of the laborious task of training the algorithms.
Deep learning will not just be for the technical specialists, however. General business managers of the future will need to understand when problems are likely to be susceptible to a deep-learning approach, and how to manage the diverse teams with more technical skills that will need to be pulled together to solve them.
There are plenty of stumbling blocks. The biggest involve data, starting with how to collect, “clean” and label it in a way that makes it useful for training machine-learning systems.
The good news is that many companies already have plenty of the raw material available. “Often, there is a lot of data already in existence and little of it gets used,” says Chui. But constant changes in the underlying information being collected means that models often need to be updated. In a third of the use cases for deep learning that McKinsey looked at, the algorithms needed to be retrained at least every month to keep them relevant.
A further challenge stems from ensuring that the data used to train a system is representative and leads to reliable answers. There is now widespread acknowledgment of the risks that come from biased data. Often, the problem stems from applying information collected for one purpose to a different problem, without making allowances for gaps in the dataset.
Most companies are at the very early stages of thinking about how to apply this form of AI in their own businesses — if they have thought about it at all. But for a future generation of managers, it could one day become a core skill.
The AI techniques that businesses will benefit the most from
The artificial neural networks used in deep learning systems are particularly suited to solving certain types of problems. Three analytical techniques have the potential to unlock the most value for businesses, according to McKinsey:
Classification. The system learns to distinguish between different items, then when confronted with a new input places it in the right category. This is the approach applied in image recognition: it could be used to identify whether products coming off a manufacturing line meet a certain visual quality standard.
Continuous Estimation. Also known as prediction, this involves estimating the next numeric value in a sequence. It could be used to forecast demand for a new product, based on information such as sales of a previous product, consumer sentiment and weather.
Clustering. An algorithm learns to create categories based on common characteristics. Presented with data about individual consumers such as their location, age and buying behaviour, for instance, an algorithm could create a set of new consumer segments.