When your boss decides to get involved with AI - a mini lesson for managers curious about Machine Learning.
Disinfecting those algorithms….
I took a job at a bricks-and-mortars business that told me they wanted “to go digital”. One of the first digital projects I identified was a machine-learning application. Reviews from the public were a big thing for this company and 40-50,000 were read each month by a group of human beings that hated the task. The team hated it so much that the company would frequently hire-in temps simply because the permanent staff couldn’t take it any more.
Of these reviews 70% of were short, positive reviews that went straight online. The remaining 30% were complex, sometimes complaints, and these were investigated by the same group of humans.
We taught a machine how to recognise the positive reviews saving the company from having humans read over 40,000 a month. As the machine learned - and the project looked like it would be a success - our corporate management felt it was time to get involved so started to give me advice - the same kind of advice Donald Trump would hand to medical experts trying to find a cure for COVID-19.
in fairness the advice handed down to me wasn’t quite injecting disinfectant but would always start “I think the algorithms should do…”. I don’t quite know why but it was the gratutious use of the plural “algorithms” that irritated the most.
Machine learning not algorithms
An algorithm tends to be an explicit, logical solution written by a human that totally understands the problem and has codified it by writing the algorithm.
Machine learning is a way of solving very complex problems but without a human having an explicit understanding of that problem, the solution or by codifying it into “the algorithms”.
A trivial example would be to give a machine, say, 100 tweets labelled as from Donald Trump and then 100 tweets labelled as from Barak Obama. Both talk about different things and use different language when tweeting. It’s not hard to imagine that given an unlabelled tweet from either men the machine would do a pretty good job of predicting if that new tweet was from Trump or Obama.
If there were algorithims involved in the predicition then their logic would be very explicit, written by a human and go something like:
if the tweet contains “make america great” or “covfefe” it’s a Trump tweet.
if the tweet contains “extraordinary” or “navigate” it’s an Obama tweet.
But machine learning isn’t explicit. In this super simple example some basic stats on the words used would allow the machine to understand if the tweet contained more Trump or Obama words.
It’s a trivial example and real machine learning applications absorb tens of thousands of examples (machines are good at that). They slice and dice the data in millions of different ways until they find an abstract solution that best matches the training they’ve been given. The machine can’t tell you why it works it just knows from its training that it’s found a solution that gives the best results.
Often there is no ‘yes/no’ answer just a degree of statistical confidence to which the machine feels its prediction is valid which is pretty much true of human prediction in real life - so much better than “the computer says no…”