Artificial intelligence (AI) algorithms have been in use for at least half a century since the term was coined. In this time, although research interest in the subject has varied, belief in its potential has remained steady. Equally, research has focused on enabling computers to glean insights from data, which can help improve the originally programmed algorithms. What is now called machine learning (ML) enables algorithms to derive new heuristics, or rules, autonomously using the data fed into the machine.
While AI and ML are often used synonymously, the latter necessarily involves a computer’s ability to use data as feedback for enhancing existing algorithms or for evolving new rules. A key feature of machine learning is the iterative process of repeatedly executing a set of tasks, incorporating the feedback on chosen data variables from the previous iteration. Even if this does not improve the algorithm, the insights from data can lead to surprising, even counterintuitive assumptions. With further training and processing more data, AI algorithms can be designed to distinguish meaningful conclusions from meaningless ones.
Making a business case for machine learning
That machine learning is now in demand has much to do with computing power becoming more affordable and more powerful than ever, allowing the collection and storage of more extensive data sets on cloud servers. In the language of computing theorists, this follows Moore’s Law, which states that the storage capacity and performance speed of computers are expected to double every year. The available computing capacity, however, needs to keep up with the data revolution. Big data – enormous sets of myriad information that arrive ever quicker – is by and large the bedrock of businesses today.
For businesses, machine learning furthers the evolutionary curve whose milestones thus far include data analytics and predictive analytics. Both of these, while involving significant amounts of data, were guided by human effort, with little or no autonomous decision-making expected from the computer. The chief advantage offered by machine learning is complexity – a more significant number of variables or “features” (to use the machine learning term), which can be factored into the analysis. This often tends to yield somewhat surprising insights from data, which in previous analyses might have only led to simple deductions.
Applications like self-driving cars offer a hint of the possibilities of gleaning insights from data and improving existing algorithms. The application of machine learning in tailoring content for streaming services is another example. The viewer’s preferences are recorded as feedback for the algorithm programmed to offer viewing suggestions. So, while straightforward analysis might suggest that a viewer who has watched one video or show classified as science fiction is likely to be interested in other shows in the same genre, machine learning enables the addition of other variables simultaneously.
How much data does a machine need?
At a higher magnitude, if the viewer were streaming these shows on a smart television, which is one of several smart devices in the home, machine learning could offer the viewer not just a selection of shows but also derive insights from data recorded by other devices in the smart home, such as a choice of food or beverage, or preferred lighting and temperature settings. In these ways, Internet of Things (IoT) systems help to realize the potential of machine learning.
All of this data offers a goldmine for the various businesses catering to the viewer described above, including manufacturers of various smart devices, Internet and mobile service providers, food and beverage retailers, and even furniture dealers. For many of these businesses, many of the questions to which they seek answers can be expressed in plain language. For instance, what would someone who enjoys watching romantic comedies usually want to drink with their dinner?
At the simplest, answering the above question requires a computer to have three data sets: data on viewers of the particular romantic comedy show, data on whether they order dinner around the time of the show, and data on what these customers pick up in the drinks aisle at their supermarket. The quality of the insights from this data is critical to answering the question specifically and sensibly. If the data is not granular enough – if it does not contain much detail, the algorithm could easily throw up spurious answers.
To put the scale of machine learning in perspective, nearly 200 million emails are sent every minute; 4.5 million YouTube videos are watched in the same duration, and nearly 5 million searches conducted on Google. Without machine learning, separating junk emails from useful ones alone would represent a staggering human effort. From a business perspective, each of those 5 million searches amounts to a marketing opportunity, provided their algorithms can record and translate the insights from that data accurately.
From data to insight, from chaos to light
The big data boom in the recent past is comparable to the oil boom or the gold rush in its power to captivate entrepreneurial imagination. Garnering insights from data using machine learning has already allowed businesses to plan their expenditure on, for instance, marketing and advertising in more nuanced ways, targeting consumers through extremely personalized approaches. The banking industry, now nearly completely networked around the globe, leverages algorithms that are self-taught to track the millions of transactions occurring every second and spot potentially fraudulent ones.
Another consequence of machine learning is the advances made in natural language processing, allowing computers to conduct human-like conversations mimicking even emotion and accent. Businesses now regularly deploy interactive response systems as the first customer interface, with customers often unaware that they are not talking to another human being. These “chatbots” are the realization of one of the oldest computing goals: replicating the (until now) uniquely human ability of speech. While Siri or an Alexa are fast becoming part of today’s households, their ability to converse like humans is still a work in progress.
The emergence of new use cases for machine learning has also led to fine-tuning the level of autonomy of machine systems. The examples discussed thus far are largely in the domain of unsupervised machine learning – i.e. the computer has to sift through mostly unstructured data to find patterns and derive explanations. Semi-supervised and supervised learning approaches are also utilized where there is a need for specific insights from data, which can be discovered more quickly with varying levels of human intervention during the execution of the algorithm.
Applications like navigation rely on the reinforcement learning approach whereby algorithms have the autonomy to iteratively discover the optimal configuration for achieving predetermined goals using programmed actions. Algorithms are tasked with processing data from the environment (i.e., traffic reports, weather conditions, etc.) to suggest, for instance, the fastest route between two points or the one with the least traffic lights. Through this process, the algorithm builds on its knowledge of routes previously plotted and learns how to utilize new data to update routes.
From machine learning to deep learning
Machine learning and big data have symbiotically contributed to furthering the possibilities of leveraging insights from data. Deep learning, which involves simultaneously deploying multiple layers of neural networks to process data, is being leveraged for a host of applications, including healthcare monitoring, advanced image processing, and complex mathematical research. The software architecture underpinning deep learning aims to duplicate human modes of thought and problem solving by examining data at multiple levels of detail in a hierarchical manner. Such multi-granularity is enabled by big data, provided that the data is verifiably accurate.
As long as sufficient computing power and processing speed — and reasonably fast network connections — are available, deep learning can be utilized wherever making data-driven decisions is hindered by a high level of data complexity. In this sense, deep learning is said to be closer to true AI, especially in learning newer ways of recognizing data patterns through processing data sequentially without human intervention. The difference from traditional machine learning approaches is that each sequence in this learning chain involves a higher degree of nonlinearity or divergence from expected input-output relationships.
Compared to even 30 years ago, there is now enough support for research into machine learning, particularly from the private sector, which is ever hungry for data. The concerns now are not so much about computing power and cost as they are about the ethics of data collection and the risks specific insights from data pose to human freedom. The use of deep learning in creating social media user profiling algorithms, for instance, may have a legitimate application in enabling security screening by law enforcement agencies. The flip side is that algorithms can be used to amplify pre-existing human biases and thereby counteract legitimate aims.