Artificial Intelligence: The Age of Machine Learning
By Kulvinder Lotay
AI is potentially the most important general-purpose technology of our era.
As the sun begins to set over the Hudson River, you glance at your watch, stepping out of the front door of Hancock. You pull out your phone as it "pings", a notification about a dinner reservation being confirmed, while you turn up the volume as your favorite after-class evening playlist comes on. Looking up, your car pulls up, you open the door and step into the back seat. A quick glance at the car heads up display shows you that Alexa, your automated home assistant, is preparing tea for your arrival. Your car begins to move, as you confirm your destination, driving you home.
Coined in 1955 by John McCarthy, the term Artificial Intelligence referred to machines that could perform tasks that required intelligence when performed by humans. Ever since, we have seen dramatic claims about future breakthroughs, setting some unrealistic expectations along the way. More recently AI has been in the headlines for its victory in what may be the world’s most complicated board game, defeating the world’s best human Go player.
The core driver of economic growth in the past centuries has been technological innovation, with general-purpose technologies such as the steam engine, electricity, and the internal combustion engine giving rise to a multitude of possibilities. AI is potentially the most important general-purpose technology of our era. More specifically, it is the area of Machine Learning (ML), the ability of a machine to improve its performance without a human explanation regarding how to accomplish the given task, ‘learning by example’. ML has become much more effective and widely available in just the past few years, with a market projected to reach $70 billion by 2020.
The reasons this matters for you are two-fold: Firstly, humans know more than we can tell, we are unable to explain our ability to recognize faces for instance. This constraint is known as Polanyi's paradox and refers to the ‘tacit’ nature of human knowledge. This has historically meant we are unable to automate many tasks. Machine Learning represents a fundamentally different approach to software development, as the machine learns from examples rather than being explicitly programmed for a particular outcome. Secondly, ML systems are mostly excellent learners and are able to achieve superhuman performance in a range of activities from diagnosing diseases to fraud detection.
What can AI do today? Over the past few years, the biggest advances in AI have been in the areas of perception, and cognition. For the former category, speech and image recognition have seen the biggest gains. For speech, think Siri, Alexa, and Google Assistant. This article you are reading was dictated to a computer, and written via speech recognition, which has seen substantial accuracy gains - an error rate of 4.9% compared to 8.5% just 1 year ago, quickly approaching the threshold where they are now equal to or better than human performance.
A recent study by Stanford computer scientists indicated that speech recognition is now approximately three times as fast as typing on a phone. Similarly, image recognition for the first time surpassed humans in terms of vision error rate; using a large database of images called ImageNet, humans err approximately 5.1% of the time, whereas ML systems now have an error rate of about 4.9%. This has opened up new possibilities, with drones and robots from Aptonomy for example being used to automate much of the work of security guards. The acceleration in gains is a result of a new approach in recent years, utilizing very large ’deep’ neural nets.
Data Indicating Vision Error Rate for Humans and Algorithms
The second category that has seen significant improvements is in cognition, including problem solving and optimization. Having beaten the finest human players in Go and poker, ML systems are now being used to further improve upon systems already optimized by human experts, such as data center cooling efficiency at Google, malware detection, and fraud detection at insurance companies.
Despite these advancements, the biggest caveat is that these ML systems are trained to do very specific tasks, and their knowledge does not generalize. This is the fallacy that an AI system’s narrow understanding implies broader understanding. The remarkable performance of AI systems in narrow areas does not translate to other related areas, indicating that we are still some ways away from machines that exhibit general intelligence.
As a result, most ML systems complement human activities, rather than replacing the entire job or process, thus making them more valuable. Think of Intelligence Augmentation, rather than Artificial Intelligence. Designing and implementing new combinations of technologies, human skills, and capital requires large-scale creativity and planning, something that machines are not very good at.
View Vision Error Rate for Humans and Algorithms Visual Data
In addition to a lack of knowledge generalization, there are some other inherent limitations of current ML systems. Due to their highly complex nature, such as in deep neural networks with millions of connections, we cannot easily identify how ML systems come to their conclusions, in an almost reverse version of Polanyi’s paradox where machines ‘know’ more than they can tell. They could, therefore, contain hidden biases, not by design, but due to the data used to train the system. Similarly, when an ML system makes an error, diagnosing and correcting the problem can be extremely difficult. However, one key takeaway is that ML systems are not designed to be perfect, but rather provide the best alternative - humans have biases and make mistakes too.
So does this mean that AI will potentially replace humans at any tasks where they achieve superhuman levels of performance? Perception and cognition cover significant ground, from being able to drive a car and making hiring decisions, to forecasting sales, and even composing music. However, one key theme underlying all of these is that it is a tool for completing tasks and problem-solving. Figuring out what problems to work on next, and leading others to tackle them will remain essential. For that reason, innovators, scientists, entrepreneurs and other creators will be some of the most rewarding positions in society in the age of ML.
In the near future, the most adaptable and flexible organizations will likely thrive, being able to respond to the new opportunities that will be presented by AI. A transformational impact in business at the scale of previous general-purpose technologies is not unforeseeable, with the potential to impact virtually every industry, transforming business models and core processes to take advantage of Machine Learning.
Photo Illustration by Marist Business Visual Data