Their project used data from the Australian Financial Security Authority (AFSA) and the Australian Tax Office (ATO) to make predictions about whether Australians who are bankrupt will comply with their obligations.
The Insolvit platform has 3 main features. The compliance dashboard provides an overview of the state of non-compliance in Australia. Our heat map visualises past non-compliance based on locations. We can then overlay other data sets to find correlations and determine why certain areas have high non-compliance. In addition, our cutting edge machine learning model allowed us to find the key attributes that lead to non-compliance.
Apart from Aaron and Denver Stove who were technical and taught coding bootcamp together, the team was a bunch of lawyers and a guy from Data 61 who was the chief legal council. “A weird mix of people”. It worked out really well because the lawyers understood the problem really well.
The team won several prizes including from AFSA which subsequently invited (and paid for) Aaron and Denver to come to come to Canberra to take the project further for them in order to implement it for real. Aaron says he was really impressed that, with so much going on in the world at the moment, the government was prepared to take a progressive stance and use the very latest technology to predict non-compliance.
The data used by the team at GovHack was just a snippet of the full set that they could access when working inside AFSA.
Aaron has always been interested in space and NASA has captured his imagination and armed with some experience with machine learning he applied for work from home Melbourne. He did not have high expectations but after a number of interviews, got a job.
When we spoke with Aaron he was working as an intern for a portfolio company of Caltech University. Caltech has a department called the Jet Propulsion Laboratory which was allowed to take some of the patents from the deep space program and apply them to industrial problems on earth. Aaron worked as a data scientist on machine learning problems, in this case, for oil and gas.
The world’s supercomputers are still not powerful enough to model things like climate change. While they are very fast, they still aren’t able to make predictions about some of these huge climate problems. One potential solution is to find hardware which is optimised for machine learning, and then do simulations of things like climate change or fluid dynamics through machine learning – basically an approximation – rather than doing the actual full simulation, the prediction will be good enough and fast.
The problem Aaron was given was to solve partial differential equations (PDEs), (all physical phenomena are described by PDEs), by approximating them rather than crunching numbers – that’s what people think will be the future of scientific computing.
One of the challenges of using AI is explaining the predictions it makes. A case based reasoning system was built that sits on top of the ML model to explain it.
The GovHack experience
Aaron enjoyed the GovHack experience. He’s worked with non-technical people a bit, but never in such a close and fast environment. The three lawyers on the team were just out of law school and were incredibly intelligent. Aaron learned a lot about how to communicate, he needed to explain to the team what was possible, and they had to explain to him in great detail what the objective was.
The hackathon was competitive but everyone was very open and friendly to each other. Teams would chat with each other to find out what they were doing.
“It’s a weird amalgamation of really smart people all focussed on the one mission of whatever they were trying to do”.
Aaron definitely recommends competing in GovHack and says you should be open to being on a team where you don’t know everyone, especially in an area that you know nothing about.
“If you told me you’re going to pair me with three lawyers I would have said… no, I don’t think so” but if that happened the project would have failed.
Aaron is returning to Melbourne and will again be teaching programming bootcamp but plans to become a machine learning engineer in Melbourne.
Article by Peter Marks for GovHack