What does an AI trainer actually do?
In this feature, we take a look inside the working lives of people whose job titles often warrant the question: 'but what do you actually do?' This week, we speak to James Ferris, an AI trainer at Impulse Screen.
What do you actually do?
The short answer is that I’m a problem solver.
The long answer is a bit more involved, but I’ll give it a go.
My work is focused on solving media and marketing problems by finding better ways to optimise digital advertising campaigns, but it is the same process for all forms of AI training.
Firstly, I have to work out exactly what the problem is to be solved, breaking it down into its essential components and defining the expected inputs and outputs. A key part of this process is figuring out if the problem can be solved simply without AI – or indeed if the problem is too complex to be solved with AI (not yet, anyway).
Next, I have to determine a suitable model for the problem. Can it be solved with a (relatively) simple machine learning implementation such as a Naive Bayes Classifier or is it something that requires the more complex world of Neural Networks – a computer system modelled on the human brain and nervous system.
The type of model used will dictate how much training data is required. Take for example, image recognition training. This can be anything from hundreds of images through to hundreds of thousands of images. As the data has to be formatted to be read properly by the chosen model this is usually the most time-consuming step.
Then it’s time to actually train the model. Basically, this means showing the model samples of the training data, over and over again, and telling it whether its output was correct or not. The idea is that the connections that help to generate the correct output are strengthened as training progresses, while those that are irrelevant are diminished. This can take some time depending on the model type, the amount of data, and the processing power available.
Finally, it’s test time. In reality, this is often done concurrently with training, allowing the development of the trained model to be observed and training stopped when performance is deemed to peak.
What does a good working day look like?
Long… finding other work to do while anxiously eyeing a set of graphs that are converging towards zero.
Ideally this would mean that the internal connections of the model have ceased to change, or are at least changing an insignificant amount, roughly indicating that the model has reached the highest level of accuracy it can achieve. Training beyond this point is pointless (and in the worst cases, actually detrimental).
What does a bad working day look like?
Long… finding other work to do while anxiously eyeing a set of graphs that are slowly approaching zero only to veer away again (which means that the AI hasn’t learned enough yet or the model is out!).
What’s the most difficult part of your job?
It isn’t necessarily the most difficult, but definitely the longest and most tedious part of my job is data collection and annotation. But there can be no short cuts. This is a crucial step, as the quality of the training dataset will directly impact the performance of the final system.
What are your KPIs?
Compared to other parts of the business, they are fairly binary. I make things work or I don’t!
What’s your favourite part about your role?
Watching a final trained model in action – it’s extremely satisfying to watch all your hard work come together as if by magic.
What’s your most stressful part of your role?
Choosing the model. There’s an enormous variety of models available, all implementing different methods or variations on methods, with varying degrees of accuracy vs. performance but only one of them is ideal for your purposes.
How does your role keep you on your toes?
AI and the world of neural networks is an extremely rapidly advancing field. It seems like every few months a new design has been developed that pushes the boundaries of performance even further.
What makes a good AI trainer?
Planning, perseverance, flexibility and a genuine interest in the field.
James Ferris is an AI trainer at Impulse Screen.
