I haven’t updated the blog in a while, as I have been very busy interviewing, going back home for a bit, and transitioning into a new role. I’m proud to say that I’m now working in Mountain View, at the Googleplex! It’s an amazing opportunity to work with some of the top machine learning and statistics researchers and practitioners in the world.
About my transition to industry: as scientists, we’re attached to the idea of the inevitable march of progress, the notion that scientific discovery will conquer all the scourges of humanity – disease, famine, global warming – and usher us into an era of pure bliss. As a vision scientist, I might say that studying neural coding will help us cure blindness in the not-so-distant future.
But the reality is that a lot of these scourges have low-tech solutions. To take the example of blindness, one of its leading causes in industrialized countries is type II diabetes, for which the causes – lack of physical activity, poor diet – are well-understood. So while I’m working on understanding hierarchical visual processing, and thinking about how that might eventually translate into a neural implant that will feed visual input directly into the brain – perhaps using nanobots – I’m missing the more obvious, immediate solution, which is to influence people’s diet and physical activity so that they don’t get diabetes in the first place.
If we truly think that one’s highest purpose is to increase the well-being of the most people – and it’s certainly a viewpoint that’s been admirably championed from the Stoics, to the Gates foundation to the recent framework of Effective Altruism – then we have to start thinking about how to measure people’s behaviour, the causal effects of different interventions, and the optimal selection of these interventions given the current evidence. This is the domain of behavioural and experimental economics, causal inference, and reinforcement learning.
So I’m very happy to be working in a group that’s been strongly influenced by the work of Hal Varian and Donald Rubin. It’s a data science group that’s particularly focused on modeling user behaviour and causal inference.
To take one example, Kay Broderson has a post over at the unofficial Google data science blog on inferring the causal impact of an intervention in the context of time-series analysis; he has some R code to get this running as well. One can view as a generalization of a difference-in-differences estimator to the scenario of more than two time points and one than one potential control.
You could use this same framework, to, say, determine the effect of the change in economic policies in post-Maoist China on sex-selective child survival, or measure how building schools in Indonesia in the late 70’s helped lift a massive number of people out of poverty. Very cool stuff. On that subject, I’m collaborating with a friend of mine on the analysis of a development project in schools in Nepal, tying all this stuff together to try to help people.
TL;DR: the statement that “academia is noble because science is intrinsically good” is dubious, and I’m glad I’m getting a chance to work in the real world.
Not to fear, I will still be leveraging my deep net/vision/machine learning wizardry for some hush-hush projects in my new work.
I’ll be writing here about some interesting papers I’m reading, along with various musings, starting later this week. The future of the blog will probably be less code-heavy – I am free from Matlab! – and will probably be more focused on machine learning than before. Until next time!