Key Points:
- The dream of creating a machine as intelligent, or more intelligent, than humans is centuries old, but progress towards this only started after the rise of computers.
- The origin of the field can be traced back to a workshop in 1956 at Dartmouth College organised by the mathematician John McCarthy. The workshop made little progress despite his optimism, but it did achieve 3 main things:
1) Named the field.
2) General goals were outlined.
3) The “Big Four” (McCarthy, Marvin Minsky, Allen Newell and Herbert Simon) met and planned for the future.
- The field is hard to define because the terms it uses (like “intelligence” are hard to define), so distinctions in vocabulary are ignored. There are two main efforts:
1) Scientific – AI researchers investigate the mechanisms of “biological” intelligence and attempt to embed it in computers.
2) Practical – The creation of programs that perform as well, or better than, humans. Whether they are actually “thinking” or not is irrelevant.
- Since 1956, there has been persistent arguments between proponents of different approaches. There is very little communication between subspecialties.
- Since the 2010s, one group of AI methods called deep learning (or deep neural networks) has become the dominant AI paradigm, despite the “anarchy of methods”. AI is defined as a broad set of approaches with the goal of creating machines with intelligence. Deep learning is one approach within the subfield of “machine learning” where machines “learn from data or their own experiences.”
- Early in the history of AI research there was a philosophical split between symbolic and subsymbolic AI.
a) Symbolic AI – A programs knowledge is in words and phrases (symbols) along with rules by which the program can combine and process these symbols to perform a task. This approach was based on the “thinking out loud” processes of students performing a task. Dominated the field for 30 years.
b) Subsymbolic AI – A program is essentially a stack of equations and does not contain human-understandable symbols. Sought to capture the unconscious thought processes of humans.
- Perceptrons (a subsymbolic tool) were invented in the late 1950s. They are an important milestone in AI research and were inspired by the way neurons process information. A perceptron has multiple numerical inputs and 1 output. The perceptron adds all of the inputs up and if it exceeds a “threshold value”, it emits a 1 (otherwise it emits a 0). A numerical weighting is multiplied to each input to simulates the brain’s “preference” for stronger connections.
- Supervised learning is a key concept in AI. The learning system is given an example and produces an output. It is then given a “supervision signal”, which tells the system how much its output is different from the correct output. The system then alters its weights and thresholds. The amount that a weighting is changed depends on its associated input value. “Stronger” inputs have weightings changed more.
- Networks of perceptrons perform well on simple perceptual tasks but it was unclear how well they would perform on more general AI tasks. The number of tasks that perceptrons could perform perfectly was very limited and perceptrons were not able to be scaled up to tasks requiring large numbers of weights and thresholds. Adding a “layer” of simulated neurons to a network of perceptrons (a multilayer neural network) increases the breadth of tasks a perceptron can perform. MNN’s now underpin much of modern AI.
- Due to the death of Frank Rosenblatt (the creator of perceptrons), the discovered limitations of perceptrons and a lack of academic funding halted research into subsymbolic approaches.
- Much of the promised AI breakthroughs have not materialised. It is hard to make AI do things that small children can do easily, but easy to make them do things like beating Chess grandmasters and diagnosing complex diseases. It exposes how little we understand of our own minds.
-
Comments