The Emergence of Artificial Intelligence
A couple of weeks ago, Google CEO Sundar Pichai told an audience at a Recode-sponsored event that for humanity, the impact of artificial intelligence could be “more profound than, I dunno, electricity or fire”. In this article, I will explore how artificial intelligence emerges from data and algorithms, and how future advances in computing will aid its development.
Data and Analytics
The term ‘big data’ describes the increasing volume, velocity, and variety of data collected by organizations. It is used as a catch-all term to describe the large data sets that organizations collect. The information in these data sets (including information about an organization’s products and services, internal processes, market conditions and competitors, supply chain, trends in consumer preferences, individual consumer preferences, and specific interactions between consumers and products, services, and online portals) can be used in either backward- or forward-looking analysis. The analytical techniques developed to analyze such data are collectively known as data analytics, and generally involve the use of computer-based quantitative models. Backward-looking methods are sometimes described as descriptive (analyzing data and developing summaries and visual depictions of important trends) or diagnostic (looking at past data to determine what went wrong). Forward-looking methods can be predictive (using past trends to predict future trends) or, at best, prescriptive (predicting future trends and suggesting organizational strategies to maximize performance according to certain measures).
Algorithms
Data analytics uses algorithms to manipulate data sets in order to extract meaningful information. An algorithm is a description of steps that are performed to complete a task. For example, you may use a series of steps to find your car in a parking lot. First, you might check to see if it is in your field of vision. Second, you might check if you have a slip or receipt that indicates where your car is. Third, you might walk down the parking lot rows in a particular methodical manner until you locate your car. If you programmed these three steps into your phone, they would become an algorithm.
Selecting the best algorithm to solve a particular problem can give an organization a significant competitive edge. An algorithm is suited to a task if it works quickly and accurately to complete that task. Google rose to dominate the web search industry based on an algorithm called PageRank (developed by Larry Page and Sergey Brin in 1996) which ranked websites based on the number (and quality) of hyperlinks to those websites from other websites. More recent examples of organizations whose business models require heavy use of a few algorithms include Uber and Stitch Fix.
Machine Learning Algorithms
While an organization can benefit greatly by using an appropriate algorithm to analyze a particular kind of data, algorithms that improve themselves as they encounter data can be exponentially more useful.
Supervised machine learning algorithms are ‘trained’ on data whose fields have been ‘labeled’. They then sort new, incoming data according to their training. For example, an algorithm may be trained to translate Vietnamese to English, to determine which inbox e-mails are spam, or to determine a car’s manufacturer when given a labeled set of the car’s attributes. The vast majority of machine learning algorithms in use today use supervised learning methods.
Other kinds of algorithms include unsupervised learning, semi-supervised learning, and reinforcement learning. Unsupervised machine learning algorithms tackle ‘unlabeled’ data directly, finding patterns in it and extracting information they consider meaningful. For example, an unsupervised machine learning algorithm may look at all the different species of plants and animals in the world, and then organize them in line with whatever attribute the algorithm considers essential. The algorithm may choose to organize them according to an attribute that makes sense to human beings (such as size, color, or mobility), or according to some other attribute that is not intuitive to us. Data scientists have also developed semi-supervised machine learning algorithms, which incorporate elements of supervised and unsupervised machine learning algorithms. Reinforcement learning reaches results through trial-and-error. Though unsupervised, semi-supervised, and reinforcement learning approaches are all still largely in the research phase, they will be applied to real-world cases soon.
Machine learning approaches can also be classified in other ways. For example, artificial intelligence scholar Pedro Domingos classifies machine learning algorithms into five schools of thought (symbolists, connectionists, evolutionaries, Bayesians, and analogizers).
One of the most promising (if perhaps currently overhyped) machine learning methods is deep learning (which falls into what Domingos refers to as the ‘connectionist’ school of thought). Deep learning uses neural networks (which consist of several layers through which information can be fed forward). Neural networks are trained through a process called backpropagation. While the actions of the nodes in a neural network are modeled on the actions of a neuron in the human brain, there are otherwise very few similarities between the workings of a neural network and that of a human brain.
Artificial Intelligence
Artificial intelligence describes machine capabilities that we expect would require human intelligence. Since this definition is subjective, whether a certain machine capability is considered artificial intelligence may change as our expectations of computers evolve. Since machines have been able to perform certain specialized tasks better than humans for centuries (for example, adding or subtracting numbers), we do not consider such tasks to require human intelligence. Some tasks which we previously assumed required human intelligence (such as translating text into a different language) can now easily be done by computers, so today we may or may not consider language translation algorithms as artificial intelligence. We still expect accurate visual perception and speech recognition to require human intelligence, so machine learning algorithms that accomplish these tasks effectively will likely be considered artificial intelligence (for the time being).
However, what really distinguishes human intelligence from machine learning algorithms is that the former is able to handle a wide variety of complex tasks. Even today’s most impressive machine learning applications, such as DeepMind’s AlphaGo computer program (which is better than any human player at the complex board game Go), are considered ‘weak’ artificial intelligence: each of them is focused only on a narrow task or set of tasks. Even large artificial intelligence projects in the pipeline, such as Toyota’s effort to improve manufacturing processes by investing $1 billion in research in the US, or Apple’s secretive self-driving car project, will likely result in ‘weak’, though impressive, artificial intelligence.
We have not yet achieved a ‘strong’ artificial intelligence (also known as Artificial General Intelligence), which can complete all the intellectual tasks that a human being can. However, the significant progress in artificial intelligence in the last few years signals that a strong artificial intelligence may be achievable in the foreseeable future.
Chips and Supercomputers
Artificial intelligence requires an appropriate machine learning algorithm, data, and computing power. Many of today’s most promising artificial intelligence technologies use neural networks. A neural network can be small and simple, or large and powerful. The larger the neural network, the more powerful it is. Reaching the limits of neural networks, and therefore of today’s artificial intelligence performance, requires using powerful computer chips. Though Intel has long been the market leader in chips for supercomputers, Nvidia, Google, and others are racing to develop chips specifically for deep learning and other artificial intelligence uses.
When it comes online later this year, the fastest computer in the world will soon be the 200 petaflop computer named Summit, at the US Department of Energy’s Oak Ridge National Lab. However, some of the fastest computers in the world do not rely on American-made chips at all. One such computer is China’s 125.4 peak petaflop supercomputer Sunway TaihuLight, which is currently the fastest supercomputer in the world. Another cutting-edge supercomputer that will come online in 2018 is the Japanese government’s 130 petaflop AI Bridging Cloud Infrastructure (ABCI). Japan’s government will rent out the use of ABCI to Japanese companies and researchers, expecting them to use it for deep learning and artificial intelligence applications.
These and other trends show that advances in artificial intelligence are both enabled by, and contribute to, hardware development in line with Moore’s law (the observation by Intel co-founder Gordon Moore in 1975 that the number of transistors on a computer chip could be expected to double every two years).
The Promise of Quantum Computing
Though Moore’s law accurately described exponential gains in computing power well into the twenty-first century, industry-leader Intel has recently taken longer to develop new chips than it has in the past. This has led some analysts to question whether chip development will continue to proceed according to Moore’s law. Intel has countered that while its chips are taking longer to develop, the capabilities of new chips once they are released are in line with the predictions of Moore’s law. Still, experts are concerned that in conventional computing, Moore’s law is hitting its ceiling and further chip miniaturization may be difficult and perhaps not economically viable.
Some experts are looking to quantum computers to develop the next set of advances in computing power. Quantum computers have been in the research phase for decades. Canadian company D-Wave is already selling what it calls quantum computers, but experts do not consider these to be true quantum computers, and the computers are likely no faster than conventional computers for most applications. Still, D-Wave computers are useful in some applications: Volkswagen used D-Wave’s quantum computing to develop an algorithm to optimize the flow of traffic in Beijing. The algorithm can be run in a few seconds using D-Wave’s computer, whereas it would take a conventional computer approximately 30 minutes to complete the same task.
Google, Microsoft, IBM, and others have indicated that they will have working quantum computers within a few years. An important milestone in the rise of quantum computing will be the achievement of ‘quantum supremacy’: the point at which quantum computers can perform certain well-defined high-performance tasks better than any classical computer can. The announcements of large technology companies notwithstanding, it is not clear when quantum computing will become commonplace: it could be anywhere from within the next few years to the next several decades.
A Transformative Technology
There has been an explosion in the creation of data, and data is the food of algorithms. Advances in computing power are enabling the operation of increasingly sophisticated and powerful machine learning algorithms. As these algorithms improve, they will match or exceed human intelligence in a number of ways. They will find application in every industry, from biotech (where they will intersect with genomics) to manufacturing. While it is not clear how long it will take for a strong artificial intelligence to emerge, a large number of weak artificial intelligences can be expected to transform human society. The impact of artificial intelligence on humanity will indeed be transformative, like that of fire or electricity before it.
Follow Lansaar Research on Medium for the latest in emerging technologies and new business models.
