May 28th, 2021
by Jeremy Young

This week I was “training” our OCR software to recognise some difficult fonts and found myself humming the 80’s pop hit “Electric Dreams”, which led me to ponder…

The software uses “neural net” technology which some people would characterise as a type of “artificial intelligence”. The training process consists of creating a “ground truth” image along with an accurate textual representation of its content. It then iterates hundreds of thousands of times adding random noise to the image and checking that it can still come up with the correct answer, at the same time altering the content of, and paths through the neural net as it progresses. Clever, but I wouldn’t say it’s very intelligent: most primary school children would be able to read the ‘noisy’ text in one iteration! 

OCR software is essentially an “expert system” in that it applies a body of accumulated static knowledge to a single, specific problem in order to produce an output. Training an OCR system is a “machine learning” process which modifies the content of the system’s static knowledge. It is largely a statistical operation and at no point displays anything remotely approaching intelligence. As with all “AI” systems, OCR is essentially a best-guess.

An “intelligent” process might be described as one which creates and executes logic which was not available to it before. The OCR training process creates data, not logic. “Genetic algorithms” can attempt to create logic by randomly generating solutions to a sub-problem and then assessing the advance that has been made towards an overall goal and selecting the best sub-solution to move forward with. At the same time, they need to be able to identify dead-ends and return to earlier solutions to try something else. 

Clever, but again, not very intelligent! A human wouldn’t try every possible option; they would sit and think about what is likely to be a profitable course to follow and prioritise efforts into things which are more likely to succeed. But how do humans know or sense what is likely to work? Archimedes came up with an understanding of how things float by relaxing in a bath. Newton came up with the theory of gravity by day-dreaming under an apple tree. No computer ever experienced a Eureka moment!

Those revelations were based on personal experiences but it is another unique ability of humankind to conceive of things outside of their experience. One of my favourite films is “Bladerunner” starring Harrison Ford which is based on the novel by Philip K Dick, “Do Androids Dream of Electric Sheep?” and which explores what it means to be human. Putting aside that even humans don’t dream of sheep but rather count them in order to fall asleep and dream of other things, the title does hint at where the gulf between machines and humans lies: having the imagination to define the problem that needs to be solved; forming the desire to want to solve it; imagining new problems which could be solved by repurposing and extending existing knowledge; these are all things that machines are nowhere near to achieving.

Another favourite film of mine, “The Terminator”, portrays a dystopian world in which a rogue network of computers assesses that humans are a threat to its existence and manufactures machines to wipe them out. Elon Musk also opines that AI is the biggest existential threat to humankind whilst at the same time investing heavily in AI projects like Deepmind and OpenAI. All we need is some off-the-wall entrepreneur to harness AI research to manufacturing capability to set us off down the road into Arnold Schwarzenegger’s world (and I don’t mean California!).

But don’t worry! We’re not there yet. When someone says they use AI, just take it with a pinch of salt.