Topics covered:

How AI began,

what does a computer really do?,

what IS AI?,

the types of AI,

some details of learning systems,

and a discussion of the benifits and dangers of AI.

TO GET A FEEL FOR WHAT AI REALLY IS, IT IS NECESSARY TO TALK ABOUT COMPUTERS AND COMPUTATION IN SOME DETAIL … SORRY!

The whole business of comptuers and hence AI began about four hunderd years ago when some Mathematicians were trying to understand if Mathematics was in some sense complete.

In other words, was it possible to know whether any math statement was either TRUE or FALSE. This question lead to the invention of computers. (how strange!!!)

.

A fellow named Alan Turing (British) was interested in the question that Leibniz posed. He decided to buld a machine that could answer Leibniz's question (was mathematics complete?)

…. a machine into which one could feed a math statement, and the machine grind away and tell whether the math statement was true or false.

HE SHOWED THAT NO SUCH MACHINE COULD BE BUILT. Quite a surprise …. the are always math statments that are undecidable … quite a bummer for the math community (LOL) but quite a boon for computation by machine.

One example of a Turing machine (shown below) is a movable paper tape upon which one can write three symbols (1,2,or 3) and can be in either of two states (A or B).

The machine operates by moving the tape one space right or left and can write 1,2, or 3 (or erase the symbol), and change its state to A or B.

An example of such a machine (it’s states and what it does given what symbol it sees on the tape) is depicted below.

YES IT IS REALLY THAT SIMPLE. THIS MACHINE CAN DO ANY COMPUTATION!!! It can execute any computer algorithm. (but Slowly).

Example of a step in a Turing machine computation …. the machine sees a 1 on the tape, is in state A, so it prints a 2, moves left one space and leaves itself in state A. Yes, this machine can execuate any computer program. Wow!!!

This is a diagram of a modern version (a stored progam computer) of Turing’s invention that can execuate any computer algorithm.

It is, like Turing’s machine, very simple. It can read and write information from an to memory, do some simple logical operations (and, or, not, plus, multiply) on what it fetches from memory and can store the result back in memory. The program counter (PC) that tells the machine what memory location to operate on next.

THE TAKE-AWAY HERE:

THE MACHINE IS VERY SIMPLE; DOES NOT HAVE ANY INTELLEGENCE OF IT’S OWN (and is fast); IT JUST DOES WHAT IT IS TOLD (PROGRAMMED TO DO).

Another diagram of a modern stored program computer. It’s instructions and data are represented by 1’s and 0’s (binary arithmetic and logic system).

In 1950 Alan Turing realized that someday computers could become so powerful (big and fast) that they might think and reason iike humans.

He developed a test for this … called the Turing test - if one is typing to a computer and one can’t tell whether it is a human or a computer … it passes the test (can think).

THE TEST WAS PASSED IN 2014 by a computer imatation a 14 year old Ukranian boy.

THIS IS THE FORMAL DEFINITION OF AI (AN INTELLEGENT AGENT).

WHEN I THINK OF AI, I IMAGINE A FANCY HOME THERMOSTAT. It has sensors (to detect the temperature in a room), actuators (to turn on the heat or cooling), some rules about how it wants the temperature in the room to be, and a set of actions (a program) to adjust the room temperaure based on what the sensors report.

IMPORTANT POINT:

THERE IS NOT ONE AI! … the data processing, mathematics, and algorithms are application specific. Doing handwriting recognition my computer is very different having a computer drive a car.

HERE WE DIVE INTO ONE SPECIAL TYPE OF AI … LEARNING SYSTEMS. It has become quite popular to model learning systems after the human brain. In particular modelling brain neurons and networks of neurons.

AN INDIVIDUAL NEURON: Inputs (numbers) come in through denrites, are summed, and that sum is checked against a threshold which determines whether the neuron ‘fires’.

A diagram of a neual net made up of individual neurons.

These nets are capable of learning by training them to produce certain outputs when given certain inputs. The key to ‘good’ learning is ‘good’ training data. This is weakness of these sytems.

Another neural net diagram …

Two examples of neural nets that learn. Google translate and handwriting recognition are done by pattern match training.

QUESTIONS: (to the audiance) Humans are 1000 times smarter than ants … someday there will be AI computers (in certain domains) that will be 1000 times smarter than humans.

Is all this AI stuff good/bad … where will it lead?