The term ‘Artificial Intelligence’ (AI) is currently trending as a revolutionary new technology that will change the face of business forever, writes Dylan Janeke, Head of Technology at Nimble Technologies. But the reality is that AI is actually not new, and the phrase is being misused to cover a broad spectrum of different concepts.
One of those is deep machine learning, which actually is new and fairly revolutionary. It’s already being applied by organisations like Amazon and Google to perform complex analytics that previously were impossible. The trouble is, for the majority of businesses, it is just too expensive and too complicated to be of much use. AI, or what many people think AI is, is unlikely to be the solution your business needs.
What AI is… and what it isn’t
AI is a generic term to describe any software that mimics human cognitive and decision-making abilities. AI has in fact been around since the first computers emerged in the 1950’s, although in a rudimentary capacity. Since its first iterations it has undoubtedly become more sophisticated, however, one of the misconceptions around it is that AI means computers can think for themselves.
Any form of AI requires data, which a computer will analyse in order to provide some form of insight, or intelligence. In typical instances a developer teaches a computer to do something by giving it specific instructions. The limit of this the developer needs to understand the problem and know how to fix it, in order to teach the computer to do the same. This has given rise to a new technique known as deep machine learning, or simply deep learning. This effectively involves providing a computer with a vast array of data and a problem to solve, and letting the computer find its own solution by analysing the data it is given.
Deep learning is a way of solving problems that humans do not understand, where computers are left to gather their own conclusions by analysing millions of examples. This deep learning is what many are simply calling AI, and it is where much of the potential for AI lies. However, while deep learning has huge possibilities, it also has huge problems that make it unusable for the vast majority of business cases.
The trouble with deep learning
Google has experimented with deep learning with DeepMind, and deep learning is what powers concepts like self-driving vehicles. Deep learning can also be used for image recognition, to determine whether images contain specific characteristics, which can be useful for blocking pornographic material or filtering other visuals.
The trouble is, to get an accurate model to train a computer, you need access to huge volumes of data, in the leagues of billions of records, so that the computer can run millions of simulations to get to the point of reaching a source of truth. This is extremely expensive and requires massive processing power and lots of time. The data is also the inherent source of problems, as Tay, one of the most infamous failures of deep learning, perfectly illustrates.
Tay was a bot introduced to Twitter by Microsoft in 2016. The idea was for it to learn from interactions with people on the social media site, by scanning posts and reading direct messages, which would be input into its training model. Tay was shut down in less than 24 hours because it became misogynistic and racist, and proved the age-old computer adage – garbage in, garbage out.
Tay is the perfect illustration of the biggest challenge of deep learning. Computers can come up with answers to anything. The problem is that the answer might not make sense because the question was wrong, or the data was insufficient or inappropriate data was used. Computers do not understand and cannot think, therefore they are unable to decide what is relevant unless they are explicitly told this. Any ‘intelligence’ relies on data, which in turn relies on a human to determine what is relevant and is subject to inherent bias.
What are the business applications?
Unfortunately, while deep learning has massive potential, it does not have a huge number of applications in everyday business because it is incredibly expensive. In most cases, businesses would be better served to try and understand the problem and then build a simple solution to solve it, rather than throwing money at deep learning to solve a problem that they do not understand. A rules-based system is infinitely more applicable than deep learning, which would typically be an overcomplicated answer to the majority of business problems.