In the day and age of Big Data, more and more data are being acquired, stored and readied for processing. Traditional means of dealing with these enormous amounts of data face significant challenges and, the same time, computing resources become more financially attractive, including cloud services, Software or Platform as a Service (SaaS/PaaS).
The rise of machine learning technology means that, for the first time, companies and individuals are able to benefit from this active field of research at a scale we’ve never seen before.
What is Machine Learning or Artificial Intelligence?
But, what exactly is “machine learning” or “artificial intelligence”? What’s behind the buzzwords? It’s well worth to look back at the beginnings of this exciting field. Arthur L. Samuel, one of the founding fathers of machine learning in 1959 described it as the “programming of a digital computer to behave in a way which, if done by humans or animals, would be described as involving the process of learning.” In a more modern phrase, one could say that machine learning algorithms are used to extract all “relevant” features of the data and apply them to make predictions about future or unknown events. What exactly “relevant” is, has to be determined for each individual use case or scenario.
Science fiction books or films often paint a distinct picture of artificial intelligence and how it interacts with humans - or sometimes wipes them out. A wide range of utopian or dystopian themes often show machines as sentient or behaving in a way humans would (or could) or even as benevolently or malevolently acting out their own agenda. Will this be reality at some point? Will there be a “SkyNet”? Maybe - or maybe not, only time will tell.
What makes machines intelligent?
More realistically, it’s better to consider the following questions: “What makes machines intelligent?” or “Can machines think?”, which are directly related to how we define and establish intelligence, not only for ourselves but also for other humans or animals.
Alan Turing OBE FRS suggested a test in 1950 to address this question, focusing on natural language conversations: human evaluator would judge a conversation between a human and a machine.
If the evaluator is not able to reliably tell the two apart, the test would be passed. A more modern version not limited to a conversation could be: “When does a machine exhibit behavior in a given field or a specific task which is equal or better than what we expect from humans?” This questions doesn’t address the development of general intelligence but focuses instead on very specific tasks a machine can do at least as well as a human, or maybe even better.
Remarkable progress has been made in the recent years. Famously, IBM’s DeepBlue computer beat the then world chess champion G. Kasparov at his own game in 1997. However, by today’s standards, DeepBlue wouldn’t get very far. Although it was able to beat a human world champion in a specific situation for the first time, its success came mainly from “brute force” calculation: Its specialized architecture contained 480 special purpose chess chips, evaluating 200 Mio positions per second and had 4,000 positions and 700,000 grandmaster games stored in a database, as well as many end-game situations. While chess is solvable by brute-force, as this game successfully demonstrated, modern problems are too complex to be solved this way.
This is what made AlphaGo’s success of playing the Chinese game of Go so significant. In March 2016, it beat Lee Sedol, a 9-dan professional Go player, in five games without handicaps. Go is much more complex than chess – too complex to be solved by brute force methods. Instead, AlphaGo combines tree search techniques with deep neural networks and was initially bootstrapped from human gameplay expertise. Once it had reached some proficiency, AlphaGo was set to play against itself using reinforcement learning methods to improve its gameplay. After it defeated the European Go champion Fan Hui, the latter helped AlphaGo improve until the final game against Lee Sedol. In contrast to DeepBlue, AlphaGo’s architecture does not rely on specialized computer chips designed for this purpose, but rather on deep learning software that can be applied to a wide range of situations with appropriate training. Google, the company behind AlphaGo, has released a comprehensive framework as open source software: TensorFlow.
Autonomous driving has been the focus of intense research in recent years as well. All major car companies are working on such systems and the field has attracted the attention of tech companies that are not traditionally seen as car manufacturers, including Google or Apple. Tesla has famously included an auto-pilot in their cars allowing autonomous driving to pilot the car in many situations already. Although so far, it’s officially labeled as an assistant, requiring the attention of the driver at all times. Apart from the technological challenge, autonomous cars face many legal and ethical hurdles before they can be omnipresent on our roads. Many of the discussions regarding autonomous driving focus on those who drive cars now. However, these advances in machine learning will give more independence to those who are currently not able to drive a car themselves: senior citizens, people with special needs – even teenagers who then no longer need to face a risky return home from a night’s adventure. It’s therefore no surprise that companies such as Uber also focus their efforts in this area: Having successfully disrupted the taxi industry, moving to autonomous vehicles is the next step, even for commercial deliveries.
Machine Learning in Retail
Machine learning is poised to disrupt retail and related industries in the near future – and early adopters are already able to draw substantial improvements by adopting machine learning technologies e.g. for replenishment and pricing. Both areas are extremely complex and depend on many factors including product details, seasons, special events, weather, past sales or the influence of other products in the assortment. Machine learning can capture the complex influence of all these factors (and many more) and derive an optimal decision on a daily basis, e.g. how many products should be ordered for the next replenishment period and/or what is the best price per product. Machine learning is ideal for such a setting as it’s not uncommon for a retailer to have tens of thousands of products for which such decisions are needed. Even experienced managers are struggling to take the best decision on a daily basis for each of the products, not alone because of the sheer quantity of decisions involved. Machine learning on the other hand can learn to make the best decision at any scale with repeatable precision.