Artificial intelligence has a long history of achieving results that would be difficult to achieve otherwise. For example, mail sorting systems using optical character recognition have been used since the 1990s. This is, after all, the source of the famous MNIST dataset of handwritten digits. The same is true for reading checks for bank deposits and assessing the creditworthiness of applicants. Financial transactions are automatically checked for fraud. This forms the backbone of many e-commerce payment systems, such as PayPal, Stripe, AliPay, WeChat, Apple, Visa, and MasterCard. Chess computer programs have been competitive for decades. Machine learning powers Internet search, recommendation, personalization and ranking. In other words, machine learning is pervasive, though often hidden from view.
It is only recently that AI has risen to prominence, primarily by solving problems that were previously considered intractable and directly affecting consumers. Many of these advances are attributed to deep learning. The following are some success stories from recent years.
Digital assistants
Smart assistants, such as Apple’s Siri, Amazon’s Alexa, and Google’s Assistant, are able to respond to voice requests with a reasonable degree of accuracy. This includes mundane jobs, such as turning on light switches, and more complex tasks, such as arranging barber appointments and offering phone assistance. This is probably the most obvious sign that artificial intelligence is affecting our lives.
A key ingredient of digital assistants is their ability to accurately recognize speech. The accuracy of these systems has gradually increased to parity with humans for some applications (Xiong et al., 2018).
Object recognition
Object recognition has also come a long way. In 2010 identifying the object in an image was quite a challenging task. In the ImageNet benchmark, researchers at NEC Labs and the University of Illinois at Urbana-Champaign achieved a top-five error rate of 28 percent (Lin et al., 2010).
In 2017, this error rate was reduced to 2.25% (Hu et al., 2018). Similarly, surprising results were achieved for bird song identification and skin cancer diagnosis.
Another indication of the progress of AI is the advent of self-driving vehicles. Although full autonomy is not yet within reach, excellent progress has been made in this direction, with companies such as Tesla, NVIDIA, and Waymo launching products that enable partial autonomy. What makes full autonomy so challenging is that proper driving requires the ability to perceive, reason, and incorporate rules into a system. Currently, deep learning is mainly used for the visual aspect of these problems. The rest is fine-tuned by engineers.
Videogames
Skill in games has been used as a yardstick for human ability. Beginning with TD-Gammon, a program for playing backgammon using reinforcement learning by time difference, algorithmic and computational advances have led to algorithms for a wide range of applications.
Compared to backgammon, chess has a much more complex space of states and set of actions. DeepBlue beat Garry Kasparov using massive parallelism, special hardware and efficient searching through the game tree (Campbell et al., 2002).
Go is even more difficult because of its huge state space. AlphaGo achieved human parity in 2015, using deep learning combined with Monte Carlo tree sampling (Silver et al., 2016).
The challenge in poker is that the space of states is large and only partially observed (we do not know our opponents’ cards). Libratus has outperformed humans in poker using efficiently structured strategies (Brown and Sandholm, 2017).
Final considerations
All this barely scratches the surface of significant applications of machine learning. For example, robotics, logistics, computational biology, particle physics, and astronomy owe some of their most impressive recent advances at least in part to machine learning, which is becoming a ubiquitous tool for engineers and scientists.
Questions have often been raised in non-technical articles about an impending artificial intelligence apocalypse and the plausibility of a singularity. The fear is that somehow machine learning systems will become sentient and make decisions, independent of their programmers, that directly impact the lives of humans. To some extent, AI already directly affects the lives of humans: creditworthiness is automatically assessed, automated drivers mostly drive vehicles, bail decisions use statistical data as input. More frivolously, we can ask Alexa to turn on the coffee machine.
Fortunately, we are far from a sentient AI system that could deliberately manipulate its human creators. First, AI systems are designed, trained, and used in specific, goal-oriented ways. Although their behavior may give the illusion of general intelligence, it is a combination of rules, heuristics and statistical patterns that underlie the design. Second, there are currently no general artificial intelligence tools that can improve themselves, reason about themselves, and modify, extend, and improve their architecture as they attempt to solve general tasks.
A much more pressing concern is how AI is used in our daily lives. It is likely that many routine tasks currently performed by humans can and will be automated. Agricultural robots will likely reduce costs for organic farmers, but they will also automate harvesting operations. This phase of the industrial revolution could have profound consequences for large sections of society, since menial jobs provide much employment in many countries. Moreover, statistical models, if applied carelessly, may lead to racial, gender, or age bias and raise reasonable concerns about procedural fairness when automated to guide consequential decisions. It is important to ensure that these algorithms are used carefully. With what we know today, this seems to us a much more pressing concern than the potential for malicious superintelligence to destroy humanity.
