Wow, this is a wild ride. I remember coming across this page because the author was from my alma mater and we were pursing the same (undergrad) degree. At the time, we could do a double major in Pure Mathematics and Statistics so long as we completed the coursework requirements, which is probably why that page even exists.
The page is ~15 years old now, and I think it should be read as though its written by a 22 yr old, more reflecting on their recent university education than a guide to how to become a working mathematician.
---
With that note, I would say if someone is eager to engage in mathematics and statistics _at an undergrad level_ (at the time at my university, it was _unusual_ for people to pursue machine learning as a major, and it was in computer science school). I would recommend really focussing on Real Analysis, and the higher statistics courses, try to find the links and the commonality between the proofs and the key ideas. I would also tell myself to not to shy away from martingale theory and link it to measure theory.
Pure mathematics is a weird world. In the moment I hated myself for choosing it in undergrad, it absolutely tanked my grades because of the weird mental state I was in. At the same time when I got to my PhD/research everything starting really started to click. It's immensely difficult to digest and consume all the content in the 12-14 odd weeks that the coursework typically demands.
As someone who isn't particularly well-versed in Bayesian "stuff". Does Bayesian non-parametric methods fall under "uninformative" + "iteration" approach?
I have a feeling I'm just totally barking up the wrong tree, but don't know where my thinking/understanding is just off.
Yeah, this one does something much less insane, i.e., it converts the paths to the tree outputs into their corresponding DNS (disjunctive normal form) and represents each term as a node (side by side in the same layer) in the NN, as described by Arunava Banerjee in "Initializing Neural Networks using Decision" [1]. The resulting NN architecture is much more reasonable than the one that treebomination produces.
Another actuary checking in from Institute of Actuaries Australia - I'm a "qualified" actuary (associate level) who decided to not even try for the fellowship exams and instead change course completely and do a PhD in Machine Learning.
No one outside of the actuarial industry cares or really knows what it means to be an actuary. Having an actuarial background in non-traditional actuarial areas is almost more of a curse than a blessing as people don't really know what to do with you. Furthermore actuaries seem to demand a premium for a cohort that don't have strong enough grounding to do ML research or enough development chops to be an ML engineer. So you end up competing with other people in the data science field...It really is a weird position to be in.
Out of curiosity, what do you think industry could/should be doing with actuaries that it currently isn't? Other than just recognizing their credentials as a signal of a strong data scientist.
h2o.ai doesn't really do data pipelines, though it does appear they are eager to go into this space through their new driverless.ai tool. However this does not appear to be open source.
