My Global Minimum

ol22
Wednesday 20 June 2018

I’ll start by introducing my research, but I don’t want to lose my good audience so I’ll keep it brief. I’m just over two weeks in and, despite some rough debugging, I have by now settled into a fairly comfortable tempo. I’ve been working with machine learning algorithms, the same kind of techniques used by the likes of Facebook and Google to target ads on the internet, and applying them to solve problems in physical chemistry. The hot topic of the day in machine learning is the deep neural network, which aims to mimic the way animals and people learn (in a much simplified way) to find patterns in vast amounts of data. The network is taught to recognise some particular feature in the data using a training algorithm.

If you think of the variables in the network as defining a landscape, with hills and valleys, then the algorithm learns by running downhill. It’s chasing after the lowest point, the so-called global minimum, where it has learned everything it possibly can from what it has been given. I picture the algorithm as a little hiker trying to find its way home, traversing a perilous path over towering mountains and gaping ravines. Every leap carries with it a risk that the numbers will suddenly explode away to infinity and the hiker will lose its way forever, but if it is too timid it can get stuck in the safety of a local minimum, lacking the courage to make the jump over the next hill. With every improvement added it’s almost heartwarming to see it develop into a confident navigator of this unpredictable terrain.That it can learn to replicate the complicated and often unintuitive results of quantum chemistry – with no guidance but the raw data – is a testament to its flexibility, and although it’s really a simple model I get a kick out of musing that it sheds light on the process of learning in general. I even wonder if there is not a (somewhat strained) comparison to leadership in there somewhere. One recurring theme in the leadership sessions seems to be that in order to improve as a leader you too must have the bravery to push yourself over obstacles when those around you might hesitate. It may be some great challenge in your work, or one as small as being the first to make a suggestion, and both can be daunting, but as you improve your techniques and add new tools to your repertoire the challenges that seemed frightening or insurmountable before can become familiar, even enjoyable.

 

I’ve really had a blast with this project so far and I’d like to thank Lord Laidlaw for making it possible as well as the CAPOD staff and speakers who run this fantastic scholarship. Your time and effort is well appreciated.

Share this story


Leave a reply

You must be logged in to post a comment.