Comment by orthoganol | Hacker News Reader

Comment by orthoganol | Hacker News Reader

Parent Full thread

Playing devil's advocate slightly, but as a web developer recently getting into ML over the last year, it's really nothing that fancy. Even the cutting edge like "word2vec using long short term memory neural networks" is still at its heart a "dumb" approach to 'intelligence'... treat the phenomena like a black box, let your system figure out the 'unknown parameters' automatically from data, don't ask too many questions. Sure there are a lot of aspects to tweak and tune, but at its heart, it's still a 'dumb' approach. It can still get a lot of mileage for a lot of types of tasks, no question, but it's pretty straightforward all in all. Today you can use very accessible Python packages or frameworks like Tensorflow to implement pretty much any of the best learning algorithm almost instantly. There are likely entire fields of ML that I am short changing, but as far as anything popular you read about, which at least I have been exposed to, it's certainly very interesting, but not particularly 'scary' in any sense.

I quickly (read everything, didn't solve the exercises, 5 days along several commutes and flights) skimmed Steinwart-Chritmann's Support Vector Machines before joining my current company in case I needed to get into more modern papers, so I could have a decent-ish background in the subject. It's a tough area, and I have an acceptable level in functional analysis.

You are mixing "applying ML black boxes" with "creating ML black boxes". The former is not specially hard (if you know the language the libraries are in, have curiosity and time), but you need to know a little bit of what you are doing unless you want to just be one of the thousand monkeys with a typewriter. The latter is hard, like very hard.

I'm not sure what you mean by "creating ML black boxes", since I don't think most people working in ML are required to invent new types of ML approaches, in the same way that Rails, Django, Sails works for most web apps. It's pretty straightforward to learn the ancillary data preparation, cross-validation, etc, and at the very least you can quickly apply a wide array of ML techniques (KNN, decision trees, LR, neural nets, etc.) with 3,4 LOC each (frameworks) and compare results, an approach you sometimes see on Kaggle. There's a reason "data science" bootcamps are putting out thousands of attendees on the job market - it's straightforward enough to get someone in the door for an interview after a few months of hard work (assuming they have either a strong programming or strong stats background).

Brief, on phone: my view of the comment addressed is that it is different doing stuff with ML that on ML, and I thought on was the addressed topic, not with

well that's not "contributing to the field", is just being a developer using libraries. The OP is talking about creating something new, which is like contributing to physics: quite hard, and requires a lot of knowledge. Maybe you don't need to be a genius (although the field attract quite a bit of them, tough competition) but surely a lot of effort.

Personally, I think the work of creating new solutions to problems by applying ML techniques is a far broader and more interesting arena than creating brand new ML techniques. Whether you're "contributing to the field" by "just being a developer using libraries" depends on what "field" you're taking about: the "field" can be "increasingly intelligent solutions to problems people have", for which successful application of "black box ML" is a very valid contribution. To continue your physics analogy: we need more people taking advantage of our knowledge of physics to engineer things like buildings, machines, and electrical circuits, than we need scientists making fundamental contributions to physics itself. Both things are useful!

Reply on news.ycombinator.com