History of Artificial Intelligence

• Creating an intelligent machine as smart as or smarter than humans is an age-old dream that became modern science with digital computers.
• The rise of computers was influenced by mathematicians' logic-based attempts to understand human thought as mechanical "symbol manipulation."
• Digital computers manipulate symbols like 0 and 1, drawing parallels between computers and the human brain, suggesting intelligence could be replicated in computer programs.
• John McCarthy organized a Dartmouth College workshop in 1956, which is often seen as the official founding of artificial intelligence.
• In 1955, John McCarthy joined Dartmouth's mathematics faculty; he had previously studied psychology and automata theory and was interested in creating a thinking machine.
• At Princeton, McCarthy met Marvin Minsky, who shared his interest in intelligent computers.
• McCarthy collaborated with Claude Shannon (information theory inventor) and Nathaniel Rochester (electrical engineer).
• McCarthy, Minsky, Shannon, and Rochester organized a 2-month study on artificial intelligence at Dartmouth in the summer of 1956.
• McCarthy invented the term "artificial intelligence" to differentiate the field from cybernetics.
• The Rockefeller Foundation received a funding proposal based on the conjecture that any aspect of learning or intelligence can be machine-simulated.
• Topics in the proposal set the agenda for the field still today: natural-language processing, neural networks, machine learning, abstract concepts, reasoning, and creativity.
• Despite computers in 1956 being much slower than smartphones, McCarthy and colleagues were optimistic about AI's progress.
• Obstacles arose such as the Rockefeller Foundation giving only half the funding requested.
• Participants had trouble agreeing.
• The Dartmouth summer of AI named the field and outlined goals, and the "big four" pioneers (McCarthy, Minsky, Allen Newell, and Herbert Simon) met.
• They left with great optimism for the field.
• McCarthy later founded the Stanford Artificial Intelligence Project with the goal of building an intelligent machine within a decade.
• Herbert Simon predicted that machines would be capable of doing any work a man can do within twenty years.
• Marvin Minsky predicted the problems of creating artificial intelligence would be solved within a generation.
• None of the predictions have materialized yet.
• It is unknown how far away we are from constructing a "fully intelligent machine".
• It is not known if reverse engineering the human brain is required, or if algorithms can produce full intelligence.

Defining Intelligence

• Voltaire's "Define your terms" relates to AI due to intelligence being ill-defined.
• Marvin Minsky's term "suitcase word" describes intelligence and related concepts like thinking, cognition, consciousness, and emotion, that are packed with different meanings.
• IQ measures human intelligence on a single scale.
• Intelligence has different dimensions: emotional, verbal, spatial, logical, artistic, social, etc.
• Intelligence can be binary, continuous, or multidimensional.
• AI mainly focuses on scientific and practical efforts, not theoretical distinctions.
• AI researchers study the mechanisms of "natural" (biological) intelligence and try to embed it in computers.
• AI proponents want to create computer programs exceeding human capabilities, regardless of whether these programs think like humans.
• Many AI people joke that their motivations depend on their funding sources.
• AI is defined as "a branch of computer science that studies the properties of intelligence by synthesizing intelligence."
• The lack of a precise AI definition has helped the field grow and advance quickly.
• Al practitioners and researchers are guided by a general sense of direction and an aim to get on with it.

Methods of AI

• At the 1956 Dartmouth workshop, participants advocated different approaches.
• Some promoted mathematical logic and deductive reasoning.
• There were the championing inductive methods using statistics from data and probabilities.
• Others believed in biology and psychology in creating brain-like programs.
• These debates persist today.
• Each approach has created its own principles, techniques, conferences, and journals, with little intercommunication.
• Since the 2010s, deep learning (or deep neural networks) has emerged as the leading AI paradigm.
• The term "artificial intelligence" itself has come to mean "deep learning", which is an unfortunate inaccuracy.
• AI includes approaches with the goal of creating machines with intelligence.
• Deep learning is one such approach.
• Deep learning is a method in machine learning, an AI subfield where machines learn from data or experiences.
• It's important to understand the philosophical split between symbolic and subsymbolic AI.

Symbolic AI

• Symbolic AI uses human-understandable words or phrases ("symbols") with rules for processing to perform tasks.
• The General Problem Solver (GPS) was an early AI program that could solve problems like the "Missionaries and Cannibals" puzzle.
• GPS’s creators, Herbert Simon and Allen Newell, recorded students "thinking out loud" while solving logic puzzles.
• GPS instructions were encoded in a symbolic manner.
• For the "Missionaries and Cannibals" puzzle, the initial state and desired state are described using symbols like LEFT-BANK, RIGHT-BANK, MISSIONARIES, CANNIBALS, and BOAT.
• The program has "operators" to transform states and "rules" to encode task constraints.
• The MOVE operator moves missionaries and cannibals and requires checks to ensure constraints like the maximum boat capacity are met.
• The computer doesn't know what the symbols mean.
• The computer's "meaning" of symbols is how they can be combined, related, and operated on.
• Symbolic AI argues that mimicking the brain is unnecessary, capturing general intelligence is possible using symbol-processing programs.
• These symbol-processing programs would use symbols, their combinations, and rules.
• Symbolic AI was dominant and formed expert systems, medical diagnoses, legal decision-making, and reasoning with common sense.

Subsymbolic AI

• Subsymbolic AI was inspired by neuroscience to capture unconscious thought processes.
• Subsymbolic AI can recognize faces.
• Subsymbolic programs are equations (hard-to-interpret operations on numbers) that learn how to perform a task.
• The perceptron, invented by psychologist Frank Rosenblatt in the late 1950s, was a brain-inspired AI program.
• Rosenblatt's perceptrons were inspired by the way neurons process information.
• A neuron receives electrical or chemical input from other neurons, sums the inputs, and fires if the sum reaches a threshold.
• Neurons have different connection strengths (synapses); stronger connections weigh more in the input sum.
• Adjusting synapse strength drives learning in the brain
• A perceptron is a computer program that simulates neuron information processing with numerical inputs and one output.
• Analogous to a neuron, the perceptron adds up its inputs and outputs either 1 (fires) or 0 (doesn't fire), depending on if the sum exceeds its threshold.
• Numerical weight is assigned to each of a perceptron's inputs, it's multiplied by its weight, then its added to the sum.
• A perceptron's threshold value is set by the programmer.
• A perceptron makes a yes-or-no decision based on whether the weighted sum meets a threshold.
• Rosenblatt proposed networks of perceptrons could perform visual tasks such as recognizing faces and objects.
• A perceptron can be designed to recognize handwritten digits by turning an image into numerical inputs and determining weights/thresholds for the correct output for each digit.
• In this, each pixel in the handwritten number becomes an input for the perceptron.
• Assigned pixel intensity from 0 to 1 serves as the numerical input.
• A perceptron is unlike the General Problem Solver.
• A perceptron lacks explicit rules for tasks, encoding knowledge into weights and threshold values.
• Rosenblatt showed a perceptron could perform perceptual tasks like recognizing digits with the correct weight and threshold values.
• A perceptron must learn values on its own.
• Rosenblatt proposed learning correct values through conditioning, similar to behavioral psychology.
• The perceptron should be trained through positive and negative reinforcement, supervised learning.

Supervised Learning

• During training, the system produces output, and then it receives a "supervision signal" describing how much its output varied from the correct one.
• The system can thus adjust its weights and thresholds
• Supervised learning requires positive examples and negative examples and each example is labeled by a human.
• The training sets trains the systens, and performance is evaluation using the test set.
• The most important term in computer science is algorithm: a "recipe" of steps a computer takes to solve a problem.
• Frank Rosenblatt's primary contribution to AI was the perceptron-learning algorithm to train the weights and thresholds.
• The algorithm initially sets random values between -1 and 1 to the weights and threshold.
• The training process compares the perceptron's output with the correct category label.
• The output is not changed if the perceptron is correct, and threshold and weights are changed making the perceptron's production closer to the correct answer.
• Higher-intensity pixels are most impactful for the amount of weight for an error.

Limitations

• The whole process is repeated for the next training example, modifying the weights and threshold a little bit each time the perceptron makes an error.
• Learning happens gradually.
• A system can settle on a set of weights and thresholds that result in correct answers is possible.
• Performance is evaluated on the test examples.
• It is possible to extend the perceptron to ten outputs (one for each digit).
• Networks of perceptrons could learn to do simple perceptual tasks, according to Rosenblatt.
• The New York Times reported ridiculously optimistic predictions as reported on a press conference Rosenblatt held in July 1958.
• The fact that a perceptron's "knowledge" is a set of is hard to reverse engineer and not symbolic, unlike the General Purpose Solver.
• It is not easy to translate these numbers into rules understandable for humans.
• Our neural firings can be considered subsymbolic.
• Advocates believe that the language-like symbols and the rules for symbol processing cannot be programmed directly.
• Neural-like architectures must be the way that intelligent symbol processing emerges from the brain.
• After the 1956 Dartmouth meeting, the symbolic camp had influence.
• Minsky believed Rosenblatt's brain-inspired approach to AI was a dead end.
• Minsky and Papert mathematically proved that perceptrons are limited.
• A perceptron augmented by adding a "layer” of simulated neurons has broader capabilities, called a multilayer neural network.
• There wasn't a general algorithm, analogous to the perceptron-learning algorithm, for learning weights and thresholds.
• Frank Rosenblatt had recognized the difficulty of training multilayer perceptrons.
• Negative speculations were part of the reason that funding for neural network research dried up in the late 1960s.
• There was a lack of government funding and so research on perceptrons and other subsymbolic AI methods largely halted.

AI Winter

• By the mid-1970s, the more general AI breakthroughs that had been promised had not materialized.
• Two reports reported negatively on progress and prospects for Al research.
• One report acknowledged that “programs written to perform in highly specialised problem domains" showed promise.
• Another concluded that the results to date were "wholly discouraging about general-purpose programs seeking to mimic the problem-solving aspects of human [brain] activity over a rather wide field
• There was a sharp decrease in government funding for Al research.
• The department of Defense drastically cut funding for basic Al research in the United States.
• It was an early example of AI having bubbles and crashes.
• Optimism occurs in the research community during phase 1.
• Results are promised, and often hyped in the news media.
• Funding pours in from government funders and venture capitalists.
• The promised don't occur in phase 2.
• Funding dries up.
• Start-up companies fold, and AI research slows.
• The "AI spring" is followed by overpromising and media hype, followed by an "AI winter."
• Field had garnered such a bad image that it was even advised not to include "artificial intelligence" on job applications.

Current Challenges

• AI was harder than people thought.
• "Easy things are hard."
• Computers that could converse with us in natural language are hard to develop.
• It is more difficult for AI to achieve than diagnosing complex diseases, beating human champions at chess and Go.
• What is easy and obvious to humans is very difficult to replicate/create with AI.
• AI has helped elucidate how complex and subtle are our own minds.

Neural Networks

• Multilayer neural networks have turned out to form the foundation of much of modern artificial intelligence.
• A network is a set of elements that are connected to one another in various ways.
• Neural networks use simulated neurons akin to the perceptrons.
• Figures show a simple multilayer neural network designed to recognize handwritten digits.
• The network has two columns (layers) of perceptron-like simulated neurons (circles).
• Units are instead of simulated neurons.
• Network has a three layer, hidden layers with 3 units and another 10 unit layer.
• Large gray arrows signify that each input has a weighted connection to each hidden unit, and each hidden unit has a weighted connection to each output unit.
• The mysterious-sounding term hidden unit simply means a non-output unit.
• The network shown has hidden and output layers.
• Multilayer network can have multiple layers of hidden units and are called deep networks.
• The "depth" of a network is simply its number of hidden layers.
• Each unit multiplies each of its inputs by the weight on that input's connection and then sums the results.
• Each unit uses its sum to compute a number between 0 and 1 (unit's "activation").
• The network performs its computations layer by layer, from left to right.
• Each hidden unit computes its activation value; these activation values become the inputs for the output units, which then compute their own activations.
• Each output unit corresponds to one of the possible digit categories.
• The activation of an output unit can be thought of as the network's confidence that it is "seeing" the corresponding digit; the digit category with the highest confidence can be taken as the network's answer -its classification.
• Multilayer neural network can learn to use its hidden units to recognize more abstract features.
• Experts use trial and error to find the best settings

Learning via Back-Propagation

• Minsky and Papert were skeptical about the learning the weights in a multilayer neural network.
• Their skepticism was largely responsible for the sharp decrease in funding for neural network research in the 1970s.
• Several groups had rebutted Minsky and Papert's speculations by developing a general learning algorithm-called back-propagation to train networks.
• Back-propagation is a way to take an error observed at the output units and "propagate" the blame for that error backward to assign proper blame to each of the weights in the network.
• This allows back-propagation to determine how much to change each weight in order to reduce the error.
• Learning in neural networks simply consists in gradually modifying the weights on connections so that each output's error gets as close to 0 as possible on all training examples.
• Back-propagation will work no matter how many units your neural network has.
• Neural networks can applied to many tasks, diverse speech recognition, stock-market prediction, language translation, and music composition.
• Connectionist networks were generally referred to as being connected to the idea that knowledge in these networks resides in weighted connections between units.
• Symbolic AI was now appearing to be brittle.
• Symbolic Al was facing another AI winter.

Approaches

• According to connectionism proponents, the key to intelligence was a brain based appropriate computational architecture.
• Rumelhart and others constructed connectionist networks (in software) as models of perception, and language development.
• In 1988, the Defense Advanced Research Projects Agency (DARPA) proclaimed neural networks were important.
• People have debated symbolic and subsymbolic approaches.
• Symbolic systems can be use human-understandable reasoning.
• Subsymbolic systems tend to be hard to interpret with no programming of complex human knowledge or logic into these systems.
• Subsymbolic systems seem better suited to perceptual or motor tasks for which humans can't easily define rules.
• Each of these approaches has had important successes in narrow areas but has serious limitations in achieving the original goals of AI.
• There have been some attempts to construct hybrid systems that integrate subsymbolic and symbolic methods, none have yet led to any striking success.

The ascent of machine learning

• Al researchers developed numerous algorithms for machine learning, leading to it becoming its own independent subdiscipline of AI.
• Machine-learning researchers disparagingly referred to symbolic AI methods as good old-fashioned AI, and roundly rejected them.
• Machine learning had its cycles of optimism, government funding, start-ups, and overpromising.
• Training neural networks and similar methods to solve real-world problems would be slow, and often did not work because of the limited computation power.
• Explosive growth of the internet would see to that.
• New AI revolution.