Ask HN: Do you work on machine learning at a large SV co and regret it? • Hacker News Board

Post by mlquestions » Sun, Jul 31, 2016, 7:01 PM UTC

I have a successful career doing regular engineering work at a large SV company. I used to work (albeit not at a super advanced level) on machine learning but it was more expedient working on tractable problems with a clear path to a solution, rather than spending a lot of time on data prep, wrangling various ML algorithms and only sometimes getting somewhere. Maybe it was me, maybe it was the tools, or a bit of both. Now that there's a huge hype cycle very well underway (especially the last 5 years) and tools are getting better (hardware and software), I'm tempted to get back into it.

Do you work on machine learning (in various incarnations incl deep learning) at Google, FB or (Microsoft, Amazon) or smaller cutting edge companies? Are you really enjoying the day to day work, or not? Why? Would you rather have built a track record in a domain involving more "linear" engineering problems with a similarly large impact? Are you a thought leader and rely on deep technical knowledge of statistics and math? Is the 'risk / reward' from a career and personal satisfaction point of view worth the time invested?

Post by PaulHoule » Sun, Jul 31, 2016, 9:22 PM UTC

The "hype cycle" for machine learning is broken in the sense that advanced machine learning is stuck in the "trough of disillusionment" yet many beginners have just leaped off the "technology trigger" That is, people's experience is not well modeled by the "hype cycle".

If there is a single example to consider it is that almost every day we see another re-implementation of the NIST digits case. Have you ever seen anyone do a similar but different problem?

Probably not. Because the vast bulk of work in a machine learning project is getting the training set, data prep, wrangling with details in the algorithms etc. It is high risk.

Because so many people are new to it they blame themselves instead of the technology so it is not so clear we are in the "trough of disillusionment"; machine learning works some of the time to create miracle products from large companies, but it is not unusual to hear that "we'll never do that again".

Future advances will depend not on algorithms so much as (i) large investments to produce training data (could we get it to work with letters instead of digits? do you take that for granted?) and (ii) finding clever ways to get the training data at less cost.