Me and AI
What is AI? Ever since people have been people they have been trying to create artificial intelligence. You might say it goes back to stories of gods creating artificial life out of clay. In more modern times people used the art of the clockmaker to make ingenious figures that could, on the hour, come out of a door on a clock and perform a little act. Mechanical automata entertained thousands of people and still do. But we wanted to create something that could recreate the mind of a human, in other words, computers that could think like people. Mechanical calculating machines culminated in Charles Babbage’s Difference Engine and Analytical Engine in the mid 19th century, but Babbage was unable to build these machines with the technology, and funding, of the time. In fact only the Difference Engine has ever been built. The main part was finished in 1991 and it was finally completed in 2002. It contains 8000 parts and weighs five tons! The Analytical Engine is even more complex and was designed to be programmed using punched cards, like those used to specify the patterns for jacquard looms.
 |
The first difference engine built, on display in the
London Science Museum
The Difference Engine was designed to calculate
tables of mathematical functions.
Photo credit: Wikipedia user: geni |
 |
| Ada Lovelace Daguerreotype by Antoine Claudet 1843 or 1850 |
Of course Babbage's machines needed programming. Ada Lovelace, daughter of the poet Lord Byron, is credited with being the first person ever, apart from Babbage himself, to write a computer program. As I understand it, she was the first to realize that computers could be used more generally than for just calculating numbers.
Perhaps my talk should be called “Me and Computers”. My first contact with computers came when I was sixteen in high school and learned how to write simple programs that could do things like find all the prime numbers less than a given number or find what day of the week any day in the Gregorian calendar fell on. It was fun! (We had no exams and no assignments.) At university programming was not part of my course. Instead I had the experience, in Applied Mathematics, of using mechanical calculators which were about the size and shape of a heavy mechanical typewriter. After you set them up you furiously turned a handle while cogs chugged and whirred until the answer popped up. The electric ones turned the handle for you and practically bounced off your desk in the process. That was fun too, for an hour or so.
In the physics lab I was able to experience using an early desktop computer, but the most interesting thing was when one of the mathematics lecturers informally introduced us to “perceptrons”. He was trying out the concept of crowd-sourcing before that was even a word!
A single perceptron is a model of a single neuron and by connecting a network of them together you can try to model the workings of a brain. Mathematically a perceptron can be viewed as a plane that divides a space of many dimensions into two. If the space is populated with the objects of interest you can use a series of these planes to box in objects of one sort and separate them from the rest. The trick is in deciding how to connect the outputs of each perceptron to the inputs of the others. You can then train your network on a set of objects reinforcing connections when you get the response you want and weakening them when you get the wrong answer. Paul, now my husband, and I worked on this together. We first of all taught our network how to play noughts and crosses. And then we were able to train it to recognize hand written examples of the letter “A” regardless whether it was written in upper or lower case. By the way, punched paper cards were used to program one of the computers we had access to. Just like Babbage and the looms.
At this stage our project was shelved as we had more important things, like exams, to spend our time on. And poor Mike, the lecturer, never got a proper report on what we had achieved.
Looking back I think we were remarkably successful. Of course the way we did it was certainly an overkill. This technique is now known as a deep neural network or DNN. It went out of favour and received very little attention for many years because people mistakenly believed it was incapable of bundling separated regions (like upper case “A” and lowercase “A”) together. Now it is a very important part of modern AI.
Schematic diagram illustrating the classification of
hand-written images of the letter "A".
In two dimensions a plane becomes a line.
When I was working in a solid state physics group in the Cavendish Laboratory in Cambridge (England) again I used a desktop computer, this time for making calculations of the intensities of x-ray diffraction patterns - classic crystallography. I also was able, with the help of the electronics workshop, to build a little microprocessor based computer to control the temperature of ovens I used to grow crystals in. This computer used the same CPU chip, a 6502, [sixty-five oh two] as the Apple II one of the first successful personal computers. I then had the fun of programming everything in machine language, which is just a series of numbers.
It was also while we were in Cambridge, in 1983, that we acquired our first personal computer. This computer included a chip that was able to analyze sound into a number of frequency bands. We were intending to try experimenting with speech recognition. But once again life had other ideas. Instead we had a baby and moved back to Australia.
There I was out of a job and in the civil service capital of the country, Canberra (which is also the real capital). There seemed to be an infinite number of computing jobs and no physics jobs for someone like me, so I decided I should do a computing course. By the time I finished that, I had had another baby and, now that I had a better idea of what it entailed, the thought of working in one of those civil service jobs seemed stultifying. Fortunately one of my lecturers offered me a job as a research assistant on her speech recognition project. This took me right back to the plans we had had when we got our first home computer. I enjoyed this work but once again we were on the move, this time to Wollongong. [Pronounced like “woollen gong”.]
Before we left Canberra my boss, Mary, took me out to lunch and offered me the opportunity to do a PhD on automatic speech recognition with her. It would have to be long distance which did turn out to be a disadvantage and cause the process to take longer than it should have but I think I would still choose to accept. The result was that I was doing most of the work, including deciding what I would work on, on my own.
Automatic speech recognition is another important part of A.I. At first I continued the type of work I had been doing in analyzing the sound waves from different samples of speech.  |
Waveform of a person saying the word "heed".
This shows air pressure versus time.
The first three segments marked on the waveform
correspond to the sounds of "h", "ee" and "d".
The final segment is the release of breath after the "d". |
 |
A spectrogram of the same sound, the word "heed".
This shows frequency versus time. In the vowel "ee"
you can see four strong frequency bands or formants.
Credit: Macquarie University Department of Linguistics |
That proved extraordinarily difficult. Too many speech sounds are almost impossible to distinguish unambiguously. This led to my second approach which was to look at the context of the sounds. What sounds can become before and after a given sound? I had access to a large corpus of data from court transcripts. I was able to work out a sort of grammar for the sounds of speech in Australian English. This is also the statistical technique that is used for completing (and correcting) text that you type into your computer or your phone. It is called a language model. It is the same idea that is used in the large language models of ChatGPT for example.
Now I still use computers to help interpret various forms of light from the heavens and to model the behavior of galaxies and the gas they are embedded in, but I don’t believe that has any relevance for AI! Never-the-less people around me need to use AI to make it possible to interpret the huge amount of data that is flooding in from missions such as the JWST. [John Webb Space Telescope]
I feel as if I have spent my working life on the edges of the sea of AI, now and again dipping my toes in and even going as far as getting ankle-deep with my work on speech recognition. AI has come a long way since the days when I was experimenting with it. It has not leapt fully formed from the ether, but it has developed steadily over many decades. It is a tool which has much to offer the world.
The thing that scares me most about AI and tools like ChatGPT is that it will become impossible to distinguish truth! It could mean the end of the internet and sharing data. Can you imagine a world without computers? Of course you ought to be able to - that was the world most of us grew up in - but I don’t know if I can.
Another thing which scares me even more, if that is possible, doesn’t relate directly to AI at all. It is that the vast computing power that quantum computing promises to deliver would make all our passwords and forms of identification simple to bypass. Nothing will be uncrackable.
Humanity has some work to do.
With those cheery thoughts I will end my talk.
Thank you Martha for delivering it!!
Susan
