What is Machine Learning, and What Can it Do?
How do we meet the challenges of an increasingly complex, fast-paced world? The answer is “we” don’t meet the challenges: humanity increasingly relies on machines and artificial intelligence, or AI, to solve our meatiest problems for us. Consider the growth of algorithms, or the propagation of systems that are designed to handle millions or even billions of new data points per day. Creating effective response and processing systems requires a machine intelligence that can understand the input it receives and help create a complicated model to generate the desired output.
If this sounds too abstract, consider the following example: an online business receives several thousand requests for information on its products per day through its website. Many of these requests may have one or more pieces of custom, desired information. In order to effectively sort the requests in a timely manner, machine learning studies the incoming emails (referred to as “input”) and then tries to develop a model that will help create responses that lead to the desired outcome, in this case an intelligent and helpful response (referred to as the “output”). This model isn’t a simple matter of sending a recipient a particular diagram or graphic; instead, the model will likely have several steps or layers that generate a complex reply suited to the initial request. Even in cases where a business licenses images, for example, the specific image must be broken down into several hidden layers of information and formulas in order for the machine to understand what it is perceiving, and what is needed.
Machine learning does this, in part, by creating a neural network. The input/output model is essentially the most basic description of such a network. Within each one of these networks there are several methods that help the machine design the layers of processing necessary to complete the task. Two of these methods are known as backward propagation and forward propagation. Backward propagation is essentially based on errors detected in the output, and works backwards to solve the issue, while forward propagation starts with the input and works through the process. In the example given above of a business that licenses images over the internet by email request, machine learning would first read the email and break down the request for the image in a series of layers. One layer would identify edges, for example, while the next layer would begin to be able to describe contours, and the next layer would start to be able to perceive shapes such as an elephant, until eventually the machine would “learn” how to correctly categorize certain shapes.
As with a human brain, each time the machine executes a task, it adds complex data to its network, which is why this process is described in terms of “learning.” Understanding how to identify an elephant versus a rhinoceros, for instance, will undoubtedly take many iterations of the process, and adjustments to both the forward and backward methods of propagation. However, the more examples the AI has, the more fluent it becomes in its work.
Obviously, businesses use this type of processing for more than sorting image requests. Companies including DeepMind Technology, Descartes Lab, Aspen Technology, H2O.ai, and numerous others design technology suited to the specific needs of their clients. From financial companies to gaming businesses to industrial waste sorting facilities, the number of applications for AI is virtually limitless. As with any complex system, the question is not can a machine be trained to process complex problems and provide intelligent, resourceful answers. Rather, the question is: how quickly can the machine do it?
Within deep machine learning, there are several distinct modes: supervised, semi-supervised, or unsupervised. Broadly speaking, the less supervision involved in the learning, the more open-ended the solution may be. As an example, supervised learning tends to be used in classification systems where the output is already a known quantity. The complexity here tends to be in translating the initial input into a suitable output, which is demonstrated in the example of the online image business above.
Unsupervised learning tends to be centered around pattern analysis, and as a result draws on commonalities witnessed in a particular set or series of data. The applications for this process are usually based on statistical analysis, and can be applied to any business or organization which is seeking to make sense of a large set of data without necessarily predicting what the results will be.
For a business such as banking, it would therefore be possible to have systems which are designed both around a supervised and unsupervised model. In the first instance, a bank may need a system that can handle a high volume of customer requests with the understanding that each request will have a known result, such as a wire transfer or the opening of an account. In an unsupervised system, the data may yield surprising patterns between the banking habits of customers who live close to a supermarket in a given geographical area versus those who must travel a long distance before encountering a similar store.
Of course, the most exciting aspect of deep machine learning and AI is not in the tasks it is currently performing, but rather the tasks it will be able to handle in the near future. The key concept here is the development of a large, possibly global neural network that possesses knowledge of all of the previous tasks and challenges each machine has had to undertake. Much like the difference between an intern and a seasoned professional with decades of experience, current learning technology is still in its early years, but is building toward a much broader and more capable level of performance and knowledge.
Speculation is of course a fool’s errand, but the possibilities for the maturation of this technology are awe-inspiring. The thorny problems humanity is currently facing, including the change in climate, are of a complexity that world governments are currently having difficulties effectively addressing. However, an experienced AI, drawing on decades of problem-solving, may be able to start to develop unimaginably complex models that take the input of a warming climate and deliver the output of a comfortable, livable planet.
