|
This is interesting. I screen a lot of ML researchers and engineers. The process typically starts with a 30-90 minute phone screen (it ends faster if I don't like you). I start by discussing the company and the position, in as much depth as the candidate wants. Then we discuss one project of the candidate's in depth, both for some technical expertise but also to suss out what the candidate, as opposed to a team, actually did. Good candidates should be able to discuss what the project did, why, implementation, challenges, problems overcome, tools used, etc. Then we do a quick survey of general ML technologies; we're hiring senior people so they should have a pretty good breadth of experiences. So they should know their way around regression, trees, forests, boosting, svm, linear algebra techniques, plus optimization tricks for all of the above. It's ok if a candidate hasn't used any one specific thing, but they should have a rough knowledge of which tool to use when, plusses and minuses of tools, and the typical ways one improves results (hold-out testing, k-folds, l1 and l2 penalties, etc etc.) If you've actually used most of these tools you should easily pass the phone screen. Then we bring people in or fly them out. Here, we go more into depth in several tools. You should know the math and the practical use of several ml techniques really well. We'll also make sure you can write enough code that you don't have to be handheld. That is, you need to be able to do some data processing and string together data pipelines to get your work done. common resume failings include (NB: this is for an ML engineer / data scientist): * (very common) no academic prep and no demonstrable experience. ie a potential junior person applying for a senior position. If he or she has no degree but spent 4 years at netflix or somewhere good, we'll definitely at minimum talk on the phone. Demonstrable experience definitely means education barely matters. * you live abroad, thus requiring sponsorship and a move, and aren't a really strong candidate * (surprisingly common) you spammed all our different jobs, making us believe you just want a job, any job, and paid no attention to what we're looking for Common failings from a phone screen or interview include: * (weird, but happens) don't make the agreed upon phone screen time without an email before (or even an email after, if an emergency occurred) * (really weird, but happens) you have no idea what the company does. I'm looking for you to have spent at least 5 minutes reading our site and have a 1 sentence summary of what the company does. If you haven't done even that much research, why on earth is the candidate applying to us? A perfectly fine example would be, for example, if I worked at google: You build a search engine and you want ML people to improve search results. At, say, square, a great response would be "you're handling monetary transactions and therefore have fraud problems from card holders and probably want to build risk management tools to protect yourself from bad merchants." I find it really weird many people don't even know that much. This isn't an instant fail, but it's a negative sign. * he or she wasted his or her phd. They may (or may not) have a very deep understanding of one tiny piece, but lack a decent understanding of the breadth of the field. We need breadth. And yes, they may be able to learn, but the question is then what the hell did they spend a 6 year phd learning? Also, we're hiring senior people. And again, anyone with a good ms in ml should do fine (not just a theory; we've hired them from stanford and ga tech). * They don't know deeply at least one or two techniques and tools. Whether its hadoop, vowpal wabbit, R, sklearn, weka, whatever, you should have become an expert in at least one tool and know it inside and out. * (very common) an inability to program enough. Obviously most ml people aren't engineers, but they have to be able to do data processing, data cleaning, run ml tools, build their own ml tools if necessary, etc. Senior people in yahoo or google's research labs may not need to program, but for us you do. * (somewhat common) bad communication skills. We speak english; you need to be able to speak reasonably clear english so that we can discuss technical topics clearly. How people can finish a long stretch in an american university without this escapes me. And to be clear, about 1/2 the employees here are foreign born, so we're pretty good with different accents and the typical indian and chinese esl idiosyncrasies. The last thing I can say about job seekers is we're talking because I have a problem. That is, I really want you to succeed if there's a good chance this will work. |
1. Hire depending on your team. There are candidates who are amazing data scientists but do retarded shit like select * from table order by rand() on a ten billion row table and bring down the cluster. Sometimes, it is OK because your team has incredibly strong engineers who can work in close combinations with these people. Sometimes you can't really afford that. At my current workplace, my manager did a really good job of hiring people who are fullstack (data + backend + viz + ml) while having strong preferences in one direction or the other. So you might have an incredibly strong ML engineer who can roll his own ML code out on Spark, while also being open to writing back-end infrastructure code once in a while. Or you might have this guy who thinks in data pipelines all day long, but will spend a bit of time thinking about mathematical model. The teams work well.
2. Hiring for ML teams is really really hard. When I was in San Francisco at my previous job, we used to screen quite a few candidates everyweek. We would see failures similar to what you posted. Either engineers who knew fuck all about ML or people who were good at the theory but didn't know how to code. Somehow, at my current job in NYC, we lucked out because we have a bunch of really awesome people who moved out of finance but want to continue to live in NYC.
> They may (or may not) have a very deep understanding of one tiny piece, but lack a decent understanding of the breadth of the field.
This is close to what I have seen from both sides of the table. But, one thing that annoys me is the number of teams where people have pet algorithms that they are obsessed about. They keep asking questions about those pet algorithms and ding a candidate for not knowing that specific algorithm in depth. One should be able to be flexible while doing this.
> They don't know deeply at least one or two techniques and tools.
I think we see people from so many different walks of life that for junior people the goal is to make sure that they know the shit that they have in their resume and drill them if they don't. As in, we don't hire for specific techniques or tools because the team has its own collection of tools/libraries (Spark/Hadoop/Python/Word2Vec/CF etc) which you will be able to pick up fast if you are good enough.