354 karma · joined July 17, 2012
It's the variations on the standard RF that have interested me. Rotation Forests, where features are partitioned randomly and rotated (by PCA or random projection) before decision boundaries are drawn. Extremely Randomized Forests, where the node splits are completely random and not based on best possible Gini/Entropy gain along some feature. There's even an interesting use of deterministic annealing out there for incorporating unlabeled data points in an attempt at semisupervised learning.
These each have their own parameters to tune, but have had slightly different performance on different problem domains. Even models require different data-- most decision trees have a quite natural way of imputing missing values, but something like a Rotation Forest can handle neither missing values nor categorical data (unless you map it to m binary features). And that complexity spooks me away from Machine Learning as a Service, where one could start failing to understand his or her models. (Plus then I'd probably be out of a job.)
I think I can safely say that asymptotic time and space concerns of a problem firmly guide much of my choices in data structure and algorithm use. Though this is maybe followed closely by aesthetic pleasure, novelty of trying something new and neat, or clarity in the functional organization/readability/maintainability of the code.
Explains and motivates the use of autoencoders (and sparsity constraints) and restricted Boltzmann machines, too.
(I would love for a practitioner to shoot this down or corroborate my hunch-- awfully new to neural nets, having basically read a monograph, played with Theano, and watched some Hinton lectures.)
There's probably a fair degree of resentment at being unable to afford a decent apartment in the city. I imagine my tune'd be different had I lucked into something livable in a cool neighborhood, rather than living in an assortment of East Bay towns and commuting by BART.
I'll be working in the peninsula soon, at one of the giants, and that will have its own joys and issues. But I'm looking forward to seeing how my view of SF evolves, as it'll soon merely be that city across the bay I visit for dinners, shows and drinks.
I also appreciated the transparency re: the retention package offered to him in the Powerset acquisition. As a younger engineer, I've spent a lot of time focusing on salary numbers (going from $70k to $90k to $100k to $122k in a few years, and across a few companies), but I'm starting to see how that's less informative than bonuses/stock grants, where the real long-term incentives and house-buying/lifestyle-company-starting capital comes from.
Are there forums where folks openly discuss numbers like these? Glassdoor can be confusing sometimes, where it's unclear if folks are putting down their pre-tax salary number or their [salary + sign-on + annual bonus + RSU] number.
It took me a good year and half to properly appreciate well-optimized, well-tested code in production, and enjoy the challenges associated there (which aren't trivial-- the neatest, ugliest NumPy script can be entirely useless at scale, or in a timeframe useful to users).
But I still personally feel flat when the knowledge gains associated with a Spike aren't as frequent. How have other developers and math-types dealt with that disappointing feeling of a plateau? Besides just changing jobs entirely, I mean.
(It should be pointed out that I was born in the mid-80s, and am frequently distracted by cat memes. But I have briefly worked for NASA.)
Also, an extension of our country on another planet would probably be the first move toward humans getting past statism entirely. But that's just a guess/hope.
But, in the end, companies like Netflix or where I work are immediately looking for cheap ways to make X happen easier, better, and cheaper. But then hopefully the smart papers go on a shelf or are easily Googleable, and the rest of us get to learn from their efforts.
I can fit long-term and short-term goals in my brain, too, mister.
In a production environment, this is probably an insane amount of transformation, feature extraction, and classification for marginally little gains in precision (as defined here). But I'm only a year or two in to building production-environment classifiers, and nothing at Netflix's scale (though not tiny either-- it is a problem if I can't do feature extraction and high-precision/-recall* classification within a few milliseconds).
* - mid 90s, for a hard NLP/social graph problem.
Only because it might compel some contrarian to build a space elevator. And I really want a space elevator.