Data Backup Digest

In one of the most impressive displays of machine learning to date, the Deep Density Displacement Model, or D3M, has created a three-dimensional simulation of our entire universe. Ironically enough, the D3M system is so complex that the original developers – a team of astrophysicists – don't even know exactly what it does or how it works.

To put it simply, D3M performs a stunningly accurate simulation – involving gravity and its effect on the universe over billions of years – within a timeframe of 30 milliseconds. While similar simulations have been performed before, they all take significantly longer to complete.

Moreover, D3M learned the algorithms and methodologies behind 8,000 other training simulations to form its own approach. It quickly began outperforming prior simulations and even started to adjust its own parameters – even though it had not yet been trained to do so.

Shirley Ho, astrophysicist with the Flatiron Institute and Carnegie Mellon University, summarized D3M's actions by saying: ''"It's like teaching image recognition software with lots of pictures of cats and dogs, but then it's able to recognize elephants. Nobody knows how it does this, and it's a great mystery to be solved."''

Advanced simulations like this aren't just fun and games – they serve a real purpose for the scientific community of the 21st century. While it's easy enough to personally observe the world around us – and we can use hardware like telescopes and satellites to see even further – there is still a limit to what we can see, reach, and explore.

This is exactly where next-gen simulations come in useful. Not only do they help scientists gain a better understanding of the past, including how our universe came into existence in the first place, but they can use these simulations to make predictions for future occurrences, too.

But traditional computer systems and simulators are notoriously slow when it comes to processing such large amounts of data. Whereas D3M is able to complete these simulations in 30 milliseconds, the average computer in use today would take 300 hours to complete with the same amount of accuracy. While you could ultimately trim this time down to a few minutes, the accuracy of the simulation would suffer tremendously.

However, there is still room for improvement. While D3M has proven itself to be far quicker and more accurate than other simulations, it still has a relative error rate of 2.8%. While it's not perfect, it's pretty close – especially when you compared it to the 9.3% relative error rate of other, less-powerful simulations.

Ho expanded on her earlier statement by saying: "We can be an interesting playground for a machine learner to use to see why this model extrapolates so well, why it extrapolates to elephants instead of just recognizing cats and dogs. It’s a two-way street between science and deep learning.”

The Deep Density Displacement Model was originally unveiled on June 24, 2019, at the Proceedings of the National Academy of Sciences. It was presented by Shirely Ho and several of her colleagues.