Roots and the Current State of AI

Original article was published by Prabishan Shrestha on Artificial Intelligence on Medium


Roots and the Current State of AI

Every now and then we hear breakthrough taking place in the field of Artificial Intelligence. In 1997, Deep Blue defeats world chess champion, Garry Kasparov, In 2011, IBM Watson wins Jeopardy, and in 2016, AlphaGo beats the World Champion, Lee Sedol, in one of the most complex games called Go. Ray Kurzweil, an American inventor, and futurist, predicts that “By 2029, computers will have emotional intelligence and be convincing as people.” In this essay I’ll try to answer the roots of AI, how some famous AI algorithms work, and what does it tell us about our own intelligence.

Roots of Artificial Intelligence

In 1950 Alan Turing published a seminal paper on the topic of Artificial Intelligence called “Computing Machinery and Intelligence” [1]. He introduced the concept of the Turing Test to the general public in this paper. Turing Test is a test of the machine’s ability to exhibit intelligent behavior indistinguishable from that of humans. Here he considers the question “Can machines think?”. He proposed the idea of “thinking machines” and gave the idea that human intelligence can be simulated in computer programs.

In 1956, a small workshop was held during summer at Dartmouth College organized by a young mathematician named John McCarthy. He approached Marvin Minsky, Nathaniel Rochester, a pioneer electrical engineer, and Claude Shannon, the inventor of Information Theory. These four trailblazers held an eight-week summer workshop and it was there John McCarthy first coined the term “Artificial Intelligence”. He later admitted that the goal was to produce genuine not “artificial” intelligence but had to be called something so called it “Artificial Intelligence”. To be distinguished from other fields such as Cellular Automata and Cybernetics, he coined this term. The proposed study, in their words, was “The study is to proceed based on the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”[2] The paper was submitted to discuss different aspects of artificial intelligence problem: Neuron Nets, Self-Improvement, Randomness and Creativity, Natural language Processing. These topics have continued to define the field to this day.
Although they were still using 800KB floppy disk, proponents of AI were optimistic that AI was in close reach despite having millions of time slower computers that of today. Though they didn’t produce anything significant that summer, they named the field and outlined the general goals of the field.

The word “intelligence” in the term “Artificial Intelligence”

The word itself is vaguely defined. We still do not know the concrete meaning of intelligence. We know that we humans are intelligent than a mouse and a mouse is intelligent than a speck of dust. The term intelligence is used in parallel with other vague terms such as thinking, consciousness, emotions. We do not even fully understand how intelligence occurs; it is said to have an emergent property [3]. An emergent property is a property that arises from the sum of individual members that do not have those properties. For example, a single ant has a very limited ability to reason and cannot accomplish a complex task but, in a colony, they accomplish surprising tasks such as building beautiful structures.

Similarly, a single neuron does not have properties like hope, pride, love, or self- awareness, yet the collection of those simple neurons tends to emerge properties like fear, pride, joy, and love. Intelligence is something that emerges from the collective property of a large number of neurons acting together. Intelligence has other dimensions such as logical, emotional, artistic, verbal, or social. One can have high emotional intelligence but low artistic intelligence. But the field of Artificial Intelligence has mostly ignored all these distinctions and has mostly focused on two efforts: Practical AI & Scientific AI. The practical side is focused on creating programs that perform tasks better than humans without worrying about whether these computer programs are thinking in a way an intelligent being thinks. On the scientific side, Researchers are focused on investigating the mechanisms of intelligent beings by simulating it in computers.

The two major philosophical Split in AI research

Symbolic AI

A symbolic approach to AI is a computer program that can process symbols (words or phrases) with a set of rules to perform a certain task. Symbolic AI does not require massive amounts of data, no training, and no probabilistic guessing. For example, to represent the phrase “John drinks water” in symbols, we write:
drinks (John, water)
Here, “John” and “water” are symbols, and “drinks” is a relation between those two symbols. In this way, we can capture information about the universe that a computer can understand.

Now to represent the above universe in symbols we write:
Initial_state:
Right_bank = [3 humans, 3 cannibals,1 boat]
Left_bank = [Empty]
Final_State:
Right_bank = [Empty]
Left_bank = [ 3 humans, 3 cannibals, 1 boat]
Here we have represented the information in the knowledge base of the program. We now apply different game playing algorithm to reach from initial to the desired state. In this way, a Symbolic AI uses a combination of symbols, and mathematical logic to do the assigned task.

Subsymbolic AI

A Subsymbolic approach to AI was inspired by the field of neuroscience and tries to capture the unconscious thought process. Here the scientists were inspired by human brains, especially how a neuron works.
In 1958, psychologist Frank Rosenblatt invented the perceptron, a neural net model that captures how the neurons process information [4].

A neuron is a nerve cell that that carries electrical impulses. Here a neuron receives electrical signals from other neurons through branching dendrites and fires an electrical signal from the Axon once a certain threshold is reached. It sums up all the electrical input from other neurons and outputs an electrical signal once a threshold is reached. Inspired by this, Rosenblatt created a neural net model that mimics this process.

Fig: a simple Perceptron

Here a perceptron simulates information similar to the neurons. In short, A Perceptron receives numerical input with its weights and outputs 1 if the sum of weights is greater than or equal to its threshold or else outputs 0. A Perceptron is a simple computer program that outputs yes or no based on sum of inputs.
Modern deep neural nets are based on this simple perceptron. We train these neural networks with millions of data and with sufficient training they are able to perform tasks such as differentiating between a cat and dog.
We hear breakthroughs taking place in the field of Artificial Intelligence which is currently dominated by deep learning. The sudden hype of AI in the last decade is largely due to big data and fast processing computers which is the main ingredient for Deep Learning.

As for state-of-the-art AI to achieve human-level intelligence is really, really far away [5].

We, humans, find this image quite depressing. We see that there is a child and a vulture. We understand that the child is deprived of food. We easily understand that the vulture is waiting for the child to die and once he dies the vulture is going to devour and eat it. Furthermore, we can deduce the human nature of the person who took this photo. We can infer what he was thinking when he took this photo.
In conclusion, this image depicts the complexity of human intelligence. It is mind-boggling that we humans can infer so much about the picture just by glancing at this 2D image for a minute. Feeding this 2D image (array of pixel values) to state-of-the-art Computer Vision can barely deduce the meaning of it. This shows that we can barely imagine how far we are in building such a machine to deduce the meaning behind this picture.

References:
Mitchell, M. (2019). Artificial intelligence: a guide for thinking humans. New York: Farrar, Straus and Giroux.

[1] A. M. TURING, I. — COMPUTING MACHINERY AND INTELLIGENCE, Mind, Volume LIX, Issue 236, October 1950, Pages 433–460, https://doi.org/10.1093/mind/LIX.236.433
[2] McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (2006). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Magazine, 27(4), 12. https://doi.org/10.1609/aimag.v27i4.1904
[3] O’Connor, T., & Wong, H. Y. (2002, September 24). Emergent Properties. Retrieved from https://plato.stanford.edu/archives/spr2012/entries/properties-emergent/
[4]Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65(6), 386–408. https://doi.org/10.1037/h0042519

[5] Karpathy, A. (n.d.). The state of Computer Vision and AI: we are really, really far away. Retrieved from http://karpathy.github.io/2012/10/22/state-of-computer-vision/
[6] How complex whole emerge from simple parts. (2018, December 20). Quanta Magazine.

[7] Newell, A., Simon, H. A., & Stanford Univ. (1995, October 1). GPS, a program that simulates human thought. Retrieved from https://dl.acm.org/doi/10.5555/216408.216432