3 ms·
Over the last few years, I've used 20% time to: a) Learn deep learning on audio with a friend, via online courses, reading research papers, and re-implementing
by sdenton4 6y ago
Over the last few years, I've used 20% time to:
a) Learn deep learning on audio with a friend, via online courses, reading research papers, and re-implementing things. Then, to put the knowledge to work, I...
b) Joined a small bioacoustics project working with external researchers to level up their ML,
c) Developed some models and deliver some results to the external researchers, and, finally,
d) Got hired onto a new team doing ML on Audio full time, largely on the strength of recommendations from bioacoustics people.
e) I've kept hacking a bit on bioacoustics, including launching the birdsong id kaggle competition earlier this year. https://www.kaggle.com/c/birdsong-recognition https://www.kaggle.com/c/birdsong-recognition
IMO, 20% time is "just marketing" until you actually put in the personal effort to make something real out of it. Doing so is non-trivial, though.
There's a real risk of falling into a 'half-ass two things' pattern. It's difficult to do exactly one day a week on some project, then cleanly drop it until the following week. Context loss is a real problem. This year, I find myself looking out for 'low stress' times during my day job to do some deeper dive on bioacoustics and create a bunch of new stuff in a kind of sprint, rather than consistently setting aside 8 hours a week. It's hard to do a research 'sprint' in two areas simultaneously; it's better to let a research question take over my brain for a while.
(I also find that my personal limit for meaningfully tracking experiment outcomes is two model architectures. I tried three at some point this year and it was kind of a disaster.)
It's tough to motivate myself to do important-but-boring things like write unit tests in 20% time, which (combined with the context shifting problem) has often lead to pain down the line.
All told, it's a hard road, but very rewarding, IMO.
- rvna 6y agoAs someone who's interested in this space, could you point me to the resources you used to ramp yourself up? Also, what are you working on w/ Audio ML?
- sdenton4 6y agoHard to point at one thing! I started with a deep learning for artists course, and adapted the stuff I saw there to an audio mini project. And then started reading the big papers in neural networks independently. I found it helpful to read the source material (eg, ml text books and academic papers), paired with blog entries (which often have a good basic idea, but skip out get wrong some details), and actually building things for my own interests. On the work side, I'm doing low bit rate speech coding, which is good fun.
- wpietri 6y agoOoh, I've been doing text-related ML for work, and have been tempted to do birdsong recognition for a hobby project over the break. Aside from that Kaggle competition, do you have any other resources you like?
- sdenton4 6y agoIf you read a bit of the kaggle discussion you'll find a lot of interesting links. The BirdClef competitions have some good write ups which, at the very least, will get you to dinner best practices for augmentation, which seems to be half the problem. Beyond the web, Nathan Pieplow has a couple great guides to birdsong, with excellent introductory essays. There's a British book called the Sound Approach to Birding with some excellent general propose knowledge, as well. If you're looking for a fun project which to my knowledge hasn't been done, 'birdsong asr' or 'phoneme' transcription would be super cool. Feel free to send me an email ($username at gmail) if you're actually digging in. :)
- wpietri 6y agoThanks! I will check these out. The notion of 'phoneme' transcription is especially interesting to me. One of the things I find unappealing about a lot of current ML is how much it seems like brute force to me. Looking for smaller structural elements seems more appealing to me.