Until the turn of the millennium, AI’s appeal lay largely in its promise to deliver, but in the last fifteen years, much of that promise has been redeemed. AI already pervades our lives. And as it becomes a central force in society, the field is now shifting from simply building systems that are intelligent to building intelligent systems that are human-aware and trustworthy.
Several factors have fueled the AI revolution. Foremost among them is the maturing of machine learning, supported in part by cloud computing resources and wide-spread, web-based data gathering. Machine learning has been propelled dramatically forward by “deep learning,” a form of adaptive artificial neural networks trained using a method called backpropagation. This leap in the performance of information processing algorithms has been accompanied by significant progress in hardware technology for basic operations such as sensing, perception, and object recognition. New platforms and markets for data-driven products, and the economic incentives to find new products and markets, have also contributed to the advent of AI-driven technology.
All these trends drive the “hot” areas of research described below. This compilation is meant simply to reflect the areas that, by one metric or another, currently receive greater attention than others. They are not necessarily more important or valuable than other ones. Indeed, some of the currently “hot” areas were less popular in past years, and it is likely that other areas will similarly re-emerge in the future.
Large-scale machine learning
Many of the basic problems in machine learning (such as supervised and unsupervised learning) are well-understood. A major focus of current efforts is to scale existing algorithms to work with extremely large data sets. For example, whereas traditional methods could afford to make several passes over the data set, modern ones are designed to make only a single pass; in some cases, only sublinear methods (those that only look at a fraction of the data) can be admitted.
The ability to successfully train convolutional neural networks has most benefited the field of computer vision, with applications such as object recognition, video labeling, activity recognition, and several variants thereof. Deep learning is also making significant inroads into other areas of perception, such as audio, speech, and natural language processing.
Whereas traditional machine learning has mostly focused on pattern mining, reinforcement learning shifts the focus to decision making, and is a technology that will help AI to advance more deeply into the realm of learning about and executing actions in the real world. It has existed for several decades as a framework for experience-driven sequential decision-making, but the methods have not found great success in practice, mainly owing to issues of representation and scaling. However, the advent of deep learning has provided reinforcement learning with a “shot in the arm.” The recent success of AlphaGo, a computer program developed by Google Deepmind that beat the human Go champion in a five-game match, was due in large part to reinforcement learning. AlphaGo was trained by initializing an automated agent with a human expert database, but was subsequently refined by playing a large number of games against itself and applying reinforcement learning.
Robotic navigation, at least in static environments, is largely solved. Current efforts consider how to train a robot to interact with the world around it in generalizable and predictable ways. A natural requirement that arises in interactive environments is manipulation, another topic of current interest. The deep learning revolution is only beginning to influence robotics, in large part because it is far more difficult to acquire the large labeled data sets that have driven other learning-based areas of AI. Reinforcement learning (see above), which obviates the requirement of labeled data, may help bridge this gap but requires systems to be able to safely explore a policy space without committing errors that harm the system itself or others. Advances in reliable machine perception, including computer vision, force, and tactile perception, much of which will be driven by machine learning, will continue to be key enablers to advancing the capabilities of robotics.
Computer vision is currently the most prominent form of machine perception. It has been the sub-area of AI most transformed by the rise of deep learning. Until just a few years ago, support vector machines were the method of choice for most visual classification tasks. But the confluence of large-scale computing, especially on GPUs, the availability of large datasets, especially via the internet, and refinements of neural network algorithms has led to dramatic improvements in performance on benchmark tasks (e.g., classification on ImageNet). For the first time, computers are able to perform some (narrowly defined) visual classification tasks better than people. Much current research is focused on automatic image and video captioning.
Natural Language Processing
Often coupled with automatic speech recognition, Natural Language Processing is another very active area of machine perception. It is quickly becoming a commodity for mainstream languages with large data sets. Google announced that 20% of current mobile queries are done by voice, and recent demonstrations have proven the possibility of real-time translation. Research is now shifting towards developing refined and capable systems that are able to interact with people through dialog, not just react to stylized requests.
Research on collaborative systems investigates models and algorithms to help develop autonomous systems that can work collaboratively with other systems and with humans. This research relies on developing formal models of collaboration, and studies the capabilities needed for systems to become effective partners. There is growing interest in applications that can utilize the complementary strengths of humans and machines—for humans to help AI systems to overcome their limitations, and for agents to augment human abilities and activities.
Crowdsourcing and human computation
Since human abilities are superior to automated methods for accomplishing many tasks, research on crowdsourcing and human computation investigates methods to augment computer systems by utilizing human intelligence to solve problems that computers alone cannot solve well. Introduced only about fifteen years ago, this research now has an established presence in AI. The best-known example of crowdsourcing is Wikipedia, a knowledge repository that is maintained and updated by netizens and that far exceeds traditionally-compiled information sources, such as encyclopedias and dictionaries, in scale and depth. Crowdsourcing focuses on devising innovative ways to harness human intelligence. Citizen science platforms energize volunteers to solve scientific problems, while paid crowdsourcing platforms such as Amazon Mechanical Turk provide automated access to human intelligence on demand. Work in this area has facilitated advances in other subfields of AI, including computer vision and NLP, by enabling large amounts of labeled training data and/or human interaction data to be collected in a short amount of time. Current research efforts explore ideal divisions of tasks between humans and machines based on their differing capabilities and costs.
Algorithmic game theory and computational social choice
New attention is being drawn to the economic and social computing dimensions of AI, including incentive structures. Distributed AI and multi-agent systems have been studied since the early 1980s, gained prominence starting in the late 1990s, and were accelerated by the internet. A natural requirement is that systems handle potentially misaligned incentives, including self-interested human participants or firms, as well as automated AI-based agents representing them. Topics receiving attention include computational mechanism design (an economic theory of incentive design, seeking incentive-compatible systems where inputs are truthfully reported), computational social choice (a theory for how to aggregate rank orders on alternatives), incentive aligned information elicitation (prediction markets, scoring rules, peer prediction) and algorithmic game theory (the equilibria of markets, network games, and parlor games such as Poker—a game where significant advances have been made in recent years through abstraction techniques and no-regret learning).
Internet of Things (IoT)
A growing body of research is devoted to the idea that a wide array of devices can be interconnected to collect and share their sensory information. Such devices can include appliances, vehicles, buildings, cameras, and other things. While it's a matter of technology and wireless networking to connect the devices, AI can process and use the resulting huge amounts of data for intelligent and useful purposes. Currently, these devices use a bewildering array of incompatible communication protocols. AI could help tame this Tower of Babel.
Traditional computers implement the von Neumann model of computing, which separates the modules for input/output, instruction-processing, and memory. With the success of deep neural networks on a wide array of tasks, manufacturers are actively pursuing alternative models of computing—especially those that are inspired by what is known about biological neural networks—with the aim of improving the hardware efficiency and robustness of computing systems. At the moment, such “neuromorphic” computers have not yet clearly demonstrated big wins, and are just beginning to become commercially viable. But it is possible that they will become commonplace (even if only as additions to their von Neumann cousins) in the near future. Deep neural networks have already created a splash in the application landscape. A larger wave may hit when these networks can be trained and executed on dedicated neuromorphic hardware, as opposed to simulated on standard von Neumann architectures, as they are today.
 Appendix I offers a short history of AI, including a description of some of the traditionally core areas of research, which have shifted over the past six decades.
 Backpropogation is an abbreviation for "backward propagation of errors,” a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent. The method calculates the gradient of a loss function with respect to all the weights in the network.
 Greg Sterling, "Google says 20% of mobile queries are voice searches," Search Engine Land, May 18, 2016, accessed August 1, 2016, http://searchengineland.com/google-reveals-20-percent-queries-voice-queries-249917.