Artificial Intelligence (AI) has received a fair amount of media hype in recent years. We regularly see articles and news reports on the latest advances in self-driving vehicles, virtual assistants, or cutting-edge scientific research aided by artificial intelligence. Likewise, there are dozens of popular TV shows and movies that render their interpretation of a future AI landscape, often not a very bright one (Westworld, Black Mirror, and Her for example). So, in a mix of overhyped expectations and actual AI technological breakthroughs, how can we get an informed view of the real achievements and prospects in the field of Artificial Intelligence? My take is to investigate the context of AI origins and the basics of how it works. In this post, I navigate through some key ideas to extract the signal from the noise.
From Edgar Allan Poe to Alan Turing
The term Artificial Intelligence was coined in the 1950s, but its essence has been around for centuries. One of my favorite examples of how the human-machine relationship has captivated our collective mind comes from the story behind Edgar Allan Poe’s 1836 essay “Maelzel’s Chess-Player”. The chess “automaton”, also called the Turk, was invented in Europe towards the end of the 18thcentury. It consisted of a mechanical man seated in a wooden cabinet where a chess board was set. Supposedly, the automaton was able to play chess without human intervention. In 1826, the machine was brought for display across the U.S. by Johann Maelzel. Many were naively convinced of its authenticity, while others including Poe, believed it was a hoax. In his lengthy essay, Poe aims to solve the mystery by presenting his analysis of the mechanical Turk’s likely “modus operandi”. Although not entirely accurate, Poe was essentially correct. As it turned out, the machine was cleverly operated by a person hidden inside the cabinet. Maelzel’s automaton was a hoax, but it illustrates the human deep-rooted drive for creating devices capable of mimicking human actions. By the way, Amazon’s Mechanical Turk service (MTurk) is named after the chess-player Turk (more on that later).
‘Can machines think?’ is a rather philosophical question that remains without a conclusive answer. This question was tackled as a thought experiment by Alan Turing back in 1950 when he proposed a method to discriminate between “actual” and “simulated” intelligence. The Turing test originally called “The Imitation Game”, requires a human judge to interrogate two subjects (a computer and a human) over a text interface so that there is no visual or auditive interaction. The machine passes the test if it can fool the judge consistently over several trials. Turing argued that if a computer’s answers are indistinguishable from a human’s, why should not we consider it to think regardless of what it is made of? Turing was not very specific about how long the test should last, what the topic should be, or who should qualify as a judge. However, Turing’s ideas helped shape AI, his famous test or at least a version of it has been used in recent years to showcase modern AI systems. Usually, these systems are chatbots designed specifically for that purpose. A quick search over the internet returns all sorts of conflicting media reports, ones that claim AI has finally passed the Turing test, while others report the opposite.
AI Singularity
This idea, also called technological singularity, is the concept of a time in the future when computers will become autonomous agents, able to do everything a human can and much more. As you would expect, there are two camps around this idea: on one hand the enthusiasts who believe it will happen within our lifetime, and on the other the skeptics who disagree altogether.
Advances in computing power in the 80s and 90s made possible the creation of computer programs that could beat humans at playing chess. In 1997 a milestone was set when IBM’s Deep Blue defeated the world chess reigning champion, Gary Kasparov. Almost 200 years after the mechanical Turk, Deep Blue legitimately achieved superiority at playing chess, but it couldn’t do anything else. More recently in 2016, DeepMind’s AlphaGo defeated the Go world champion Lee Sedol. It is worth noting that due to all the possible board configurations, the game of Go is far more complex than chess. But still, AlphaGo is only good at playing Go. This type of intelligence where the computer focuses on a single task is called “narrow AI” or “weak AI”. Alternatively, the sort of AI that we see portrayed in TV shows such as Westworld, where robots take over humanity (…spoiler alert?), is called Artificial General Intelligence (AGI) or “strong AI”. Sci-Fi literature and cinema often associate AI with humans’ worst qualities. Power-lust and greedy machines are presented as the inevitable future of AI. But arguably, a more plausible scenario does not end in humanity’s submission to evil robots. Instead, we are more likely to see a deeper integration between humans and our intelligence-enhancing technologies, a sort of symbiotic relationship. It may seem like a theme more appropriate for a work of fiction, but AGI is an active research subject.
While the latest achievements in AI technologies (narrow AI) have been spectacular, especially in deep learning, reaching AGI is far from realized. For now, it seems safe to bet that the singularity won’t be happening any time soon. Needless to say, trying to pin down a date when AGI will be a reality is a rather speculative matter.
Machines that Learn
The formal field of Artificial Intelligence can be subdivided into an array of subfields or approaches. Out of the pool, the most popular of such approaches today is machine learning (ML).
In “classical” programming, a set of rules (a program) and data are provided as inputs by humans. The data is then processed according to the given rules, and out comes an answer. This model is known as “Symbolic AI” and was the dominant approach in the early years of AI. In contrast, machine learning models are “trained” to find out the rules, then these rules are applied to new data to obtain new answers.
Supervised ML models “learn” by looking for statistical patterns in labeled examples (lots of examples). There are two other types of machine learning models that do not need labeled data, these are: Unsupervised models and Reinforcement Learning models.
Consider an image classification system, where you want to automate the task of tagging a set of pictures. To accomplish this task, you could present to the model a set of pictures previously labeled (by humans), during this training phase the model would learn the statistical rules for associating particular pictures (inputs) to particular tags (targets). The same method is used in speech recognition systems, in this case, the examples could be audio files of people speaking.
An algorithm is defined as a set of rules or steps to follow to perform a task. In ML a crucial step is to measure how the algorithm is performing. This is done by evaluating the algorithm’s current output against the expected output. The measurement is then used as feedback to adjust how the algorithm works, and this process is what we call learning. Remarkably, this idea is useful for solving a wide variety of tasks, from recognizing dogs in pictures to autonomous vehicle driving.
The machine learning “toolbox” encompasses the subfield of deep learning, a method that works with models called neural networks. A neural network is constructed by layers of connected “neurons”: an input layer, the hidden layers, and an output layer. Neurons in each layer provide inputs and outputs for adjacent layers, and learning is done by gradually adjusting the difference between the current and expected output. This is an oversimplification, but it illustrates the basic idea of neural networks. A misconception that we often see in popular science articles is that neural networks work like biological brains. While deep learning models were developed borrowing inspiration from neurobiology, neural networks are neither brain models nor operate like a brain. ML systems in general work with statistical methods that map inputs to targets (labels) and this mapping is done by exposing the model to many examples of inputs and targets.
While AI models are capable of “learning” on their own and automating a wide variety of intellectual tasks, there are still things that need to be done by humans. Leveraging this fact, Amazon offers its Mechanical Turk service. It is a crowdsourcing service that let businesses hire “crowd workers” to perform Human Intelligence Tasks or HITs. For example, manually tagging thousands of images that would later be used to train ML models. Unlike the original mechanical Turk, Amazon is not trying to fool anyone, according to their website:
“While technology continues to improve, there are still many things that human beings can do much more effectively than computers, such as moderating content, performing data deduplication, or research.”
Such a statement coming from one of the “tech giants” tells a lot about the true status of current AI capabilities.
Without a doubt, AI will continue to improve as it becomes more present in many aspects of society and our everyday routines. However, short-term expectations for AI technologies usually tend to be overstated. With today’s abundance of media outlets reporting on the subject, perhaps it is sensible to remain a bit skeptical if we want to avoid being fooled by the “mechanical Turks” out there.
( This article was originally published by the author on September 2022 at randomforest.cc )
References and further reading:
- Poe, E. A. (1836, april). Maelzel’s Chess-Player. Southern Literary Messenger. https://www.eapoe.org/works/essays/maelzel.htm
- Eschner, K. (2017). Debunking the Mechanical Turk Helped Set Edgar Allan Poe on the Path to Mystery Writing. Smithsonian Magazine. https://bit.ly/3Dgr7RU
- Turnig, A. M. (1950, october). Computer Machinery and Intelligence. Mind, 59(236). https://doi.org/10.1093/mind/LIX.236.433
- Mitchell, M. (2019). Artificial Intelligence: A Guide for Thinking Humans. Farrar, Straus and Giroux.
- Chollet, F. (2018). Deep Learning with Python. Manning Publications Co.
- DeepMind. (n.d.). AlphaGo. https://www.deepmind.com/research/highlighted-research/alphago
- Amazon. (n.d.). Amazon Mechanical Turk. https://www.mturk.com
