Written by: Team PredictiveHire
How Victoria defeated COVID with individual action and data
Action and data
COVID has taught us that on reflection the focus on individual action with a community benefit as a goal is really a focus that leads to the greater good. In our home state of Victoria, Australia now 7 straight days with ZERO new cases. It has been an effort founded on facts and science over misinformation. In Victoria, many sacrificed a lot for their well-being for ALL. If anything, there is now proof, thanks to Victorians, that when we see facts, listen to science and let data show you how to lead that change, you can make it happen.
We’re using this approach to build a new vision for inclusive hiring.
AI, especially predictive machine learning models, are an outcome of a scientific process, it’s no different to any other scientific theory, where a hypothesis is being tested using data.
The beauty of the scientific method is that every scientific theory needs to be falsifiable, a condition first brought to light by the philosopher of science Karl Popper. In other words, a theory has to have the capacity to be contradicted with evidence.
It is how science is able to progress without bias.
There are three decisions that are made by a human in building that scientific experiment.
- Forming a meaningful hypothesis
- Data collection methodology (experiment set up)
- The data you rely on to test the hypothesis
One can argue 2 and 3 are the same as if the methodology is not sound the data collection wouldn’t be either. That’s why there is so much challenge and curiosity as there should be about the data that goes into an algorithm.
Think of an analogy in a different field of science: the science of climate change.
A scientist comes up with a hypothesis that certain factors drive an increase in objective measures of climate warming, eg CO2 emissions, cars on the road, etc. That’s a hypothesis and then she tests it using statistical analysis to prove or disprove that her hypothesis holds beyond random chance.
The best way to make sure you are following a sound scientific approach is to share your findings with the broader scientific community. In other words, publish in peer-reviewed mediums such as journals or conferences so that you are open to scrutiny and arguments against your findings.
Or to put it another way, be open for your hypothesis to be falsified.
In AI especially, it is also important to keep testing whether your hypothesis holds over time as new data may show patterns that lead to disproving your initial hypothesis. This can be due to limitations in your initial dataset or assumptions made that are no longer valid. For example, assuming the only information in a resume related to gender are name and explicit mention of gender or a certain predictive pattern such as detecting facial expressions are consistent across race or gender groups. Both of these have been proven wrong*.
The only way to improve our ability to predict, be it climate change or employee performance, is to start applying the scientific method and be open to adjusting your models to better explain new evidence.
Therefore the idea that a human can encode their own biases in the AI — well it’s just not true if the right science is followed.
* Amazon scraps secret AI recruiting tool that showed bias against women (https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G)
* Researchers find evidence of bias in facial expression data sets (https://venturebeat.com/2020/07/24/researchers-find-evidence-of-bias-in-facial-expression-data-sets/)