A Brief History of Artificial Intelligence: The First AI Winter
A Brief History of Artificial Intelligence: The First AI Winter
In 1950, a man named Alan Turing wrote a paper suggesting how to test a “thinking” machine. He
believed if a machine could carry on a conversation by way of a teleprinter, imitating a human with no
noticeable differences, the machine could be described as thinking. His paper was followed in 1952 by
the Hodgkin-Huxley model of the brain as neurons forming an electrical network, with individual
neurons firing in all-or-nothing (on/off) pulses. These combined events, discussed at a conference
sponsored by Dartmouth College in 1956, helped to spark the concept of artificial intelligence.
“Google Assistant brings together all of the technology and smarts we’ve been building for years, from
the knowledge graph to natural language processing. Users can have a natural conversation with Google
to help them in their user journeys.”
The development of AI has been far from streamlined and efficient. Starting as an exciting, imaginative
concept in 1956, artificial intelligence research funding was cut in the 1970s, after several reports
criticized a lack of progress. Efforts to imitate the human brain, called “neural networks,” were
experimented with, and dropped.
Their most advanced programs were only able to handle simplistic problems, and were described as toys
by the unimpressed. AI researchers had been overly optimistic in establishing their goals (a recurring
theme), and had made naive assumptions about the difficulties they would encounter. After the results
they promised never materialized, it should come as no surprise their funding was cut.
The stretch of time between 1974 and 1980 has become known as ‘The First AI Winter.’ AI researchers
had two very basic limitations — not enough memory, and processing speeds that would seem abysmal
by today’s standards. Much like gravity research at the time, Artificial intelligence research had its
government funding cut, and interest dropped off. However, unlike gravity, AI research resumed in the
1980s, with the U.S. and Britain providing funding to compete with Japan’s new “fifth generation”
computer project, and their goal of becoming the world leader in computer technology.
The First AI Winter ended with the promising introduction of “Expert Systems,” which were developed
and quickly adopted by large competitive corporations all around the world. The primary focus of AI
research was now on the theme of accumulating knowledge from various experts, and sharing that
knowledge with its users. AI also benefited from the revival of Connectionism in the 1980s.
Expert Systems
Expert Systems were an approach in artificial intelligence research that became popular throughout the
1970s. An Expert System uses the knowledge of experts to create a program. The process involves a user
asking the Expert System a question, and receiving an answer, which may or may not be useful. The
system answers questions and solves problems within a clearly defined arena of knowledge, and uses
“rules” of logic.
The software uses a simplistic design and is reasonably easy to design, build, and modify. Bank loan
screening programs provide a good example of an Expert System from the early 1980s, but there were
also medical and sales applications using Expert Systems. Generally speaking, these simple programs
became quite useful, and started saving businesses large amounts of money. (Expert systems are still
available, but much less popular.)
The AI field experienced another major winter from 1987 to 1993. This second slowdown in AI research
coincided with XCON, and other early Expert System computers, being seen as slow and clumsy. Desktop
computers were becoming very popular and displacing the older, bulkier, much less user-friendly
computer banks.
Eventually, Expert Systems simply became too expensive to maintain, when compared to desktop
computers. Expert Systems were difficult to update, and could not “learn.” These were problems desktop
computers did not have. At about the same time, DARPA(Defense Advanced Research Projects Agency)
concluded AI “would not be” the next wave and redirected its funds to projects more likely to provide
quick results. As a consequence, in the late 1980s, funding for AI research was cut deeply, creating the
Second AI Winter.
Natural language processing (NLP) is a subdivision of artificial intelligence which makes human language
understandable to computers and machines. Natural language processing was sparked initially by efforts
to use computers as translators for the Russian and English languages, in the early 1960s. These efforts
led to thoughts of computers that could understand a human language. Efforts to turn those thoughts
into a reality were generally unsuccessful, and by 1966, “many” had given up on the idea, completely.
During the late 1980s, Natural language processing experienced a leap in evolution, as a result of both a
steady increase in computational power, and the use of new machine learning algorithms. These new
algorithms focused primarily on statistical models – as opposed to models like decision trees. During the
1990s, statistical models for NLP rose dramatically.
Intelligent Agents
In the early 1990s, artificial intelligence research shifted its focus to something called intelligent agents.
These intelligent agents can be used for news retrieval services, online shopping, and browsing the web.
Intelligent agents are also sometimes called agents or bots. With the use of Big Data programs, they
have gradually evolved into digital virtual assistants, and chatbots.
Machine Learning
Machine learning is a subdivision of artificial intelligence and is used to develop NLP. Although it has
become its own separate industry, performing tasks such as answering phone calls and providing a
limited range of appropriate responses, it is still used as a building block for AI. Machine learning, and
deep learning, have become important aspects of artificial intelligence.
• Boosting: In 1990, Robert Schapire introduced the concept of boosting in a 1990 paper, The
Strength of Weak Learnability. Schapire wrote, “A set of weak learners can create a single strong
learner.” The majority of boosting algorithms are repetitive weak learning classifiers that, when added
together, form a strong classifier.
• Speech Recognition: Most of the speech recognition training being done is the result of a deep
learning technique referred to as long short-term memory (LSTM). This is based on a neural
network model developed in 1997, by S. Hochreiter and Jürgen Schmidhuber. The LSTM
technique supports learning tasks which use memories of thousands of small steps (this is
important for learning speech). Around 2007, LSTM began surpassing the more established
speech recognition programs. During 2015, Google’s speech recognition program reported a 49
percent increase in performance by using a LSTM that was CTC-trained.
• Facial Recognition: In 2006, the National Institute of Standards and Technology sponsored the
“Face Recognition Grand Challenge,” and tested popular facial recognitionalgorithms. Various iris
images, 3D face scans, and high-resolution facial images were examined. They found some of the
new algorithms to be ten times as accurate as the facial recognition algorithms popular in 2002.
Some of the new algorithms could surpass humans in recognizing faces (these algorithms could
even identify identical twins). In 2012, an ML algorithm developed by Google’s X Lab could sort
through and find videos which contained cats. In 2014, the DeepFace algorithm was developed
by Facebook — it recognized people in photographs with the same accuracy as humans.
Digital virtual assistants understand spoken commands, and respond by completing tasks.
In 2011, Siri (of Apple) developed a reputation as one of the most popular and successful digital virtual
assistants supporting natural language processing. Online assistants, such as Alexa, Siri, and Google, may
have started as convenient sources of information about the weather, the latest news, and traffic
reports, but advances in NLP and access to massive amounts of data have transformed digital virtual
assistants into a useful customer service tool. They are now capable of doing many of the same tasks a
human assistant can. They can even tell jokes.
Digital virtual assistants can now manage schedules, make phone calls, take dictation, and read emails
aloud. There are many virtual digital assistants on the market today, with Apple’s Siri, Amazon’s Alexa,
Google Assistant, and Microsoft’s Cortana as well known examples. Because these AI assistants respond
to verbal commands, they can be used hands-free, allowing a person to drink their coffee, or change
diapers, while the assistant accomplishes the assigned task.
These virtual assistants represent the future of AI research. They are driving cars, taking the form of
robots to provide physical help, and performing research to help with making business decisions.
Artificial intelligence is still evolving and finding new uses.
Chatbots and digital virtual assistants are quite similar. Chatbots (sometimes called “conversational
agents”) can talk to real people, and are often used for marketing, sales, and customer service. They are
typically designed to have human-like conversations with customers, but have also been used for a
variety of other purposes. Chatbots are often used by businesses to communicate with customers (or
potential customers) and to offer assistance around the clock. They normally have a limited range of
topics, focused on a business’ services or products.
Chatbots have enough intelligence to sense context within a conversation and provide the appropriate
response. Chatbots, however, cannot seek out answers to queries outside of their topic range or perform
tasks on their own. (Virtual assistants can crawl through the available resources and help with a broad
range of requests.)
In my humble opinion, digital virtual assistants and chatbots have passed Alan Turing’s test, and achieved
true artificial intelligence. Current artificial intelligence, with its ability to make decisions, can be
described as capable of thinking. If these entities were communicating with a user by way of a teletype,
a person might very well assume there was a human at the other end. That these entities can
communicate verbally, and recognize faces and other images, far surpasses Turing’s expectations.