Machine Learning: What Is It? | Justin Wang

Throughout the last decade, computers have transformed the world we live in, and have inundated our lives with technological aspects. This machine revolution has extended to our information through services like Google and other search engines. However, the information we search for is collected in other ways as well – we are nearly constantly monitored (even at an anonymous level) for our data and searches. How do these companies know our trends, the things we search for, and how to target them? Machine learning. Relatively recent advances in programming led to the development of machines that can “learn”, although not at the capacity that humans can. Before you fret about our new artificial overlords, know that machine learning and AI have one important distinction – scope. AI, in its most primitive form, is still a long way from being self-sufficient, due to the scope that it must cover to even come close to an interactable form with humans. Machine learning, which is often focused on one specific concept, is much more prevalent and is deployed in multiple applications across the machine-scape (https://quickdraw.withgoogle.com/ is a perfect example of this). Machine learning functions primarily on examples. Programmers code in “nodes”, or certain instructions which can be “weighted”, much like the categories in a gradebook. Then, the programmer gives the robot an answer key to a large amount of examples, and asks the robot to form its own weights on the nodes to create a custom-made program. For example, if a robot was asked to distinguish a bee from a three, the robot would start with a large amount of bee and three example photos. Then, the robot would draw conclusions on those example photos, test itself, and edit the weights accordingly. This cycle repeats over and over again, until the weights are close to a 99% or higher success rate. This is the exact same process that is used in advertisement targeting, where search terms are often analyzed by adaptive machine learning algorithms to predict the interests of the account searching. However, an important fact of this process is that when the programmer looks at the code after the process has been completed, the weighting that took place within the process is completely unrecognizable to whoever created the initial learning program. In other words, once the program is completed, there is not a single person on the Earth that knows exactly how the machine learning program distinguishes a three from a bee.