This post is a little bit out of place. I wrote it around the time I was starting to cover a specific topic in each entry, so I didn't want to do an "article of the day" post like Day Ten (Not Hotdog). But I didn't have all the pieces I needed to cover another aspect of neural network architecture either, and I was considering folding it into Day Fifteen (which I already knew was going to be about finetuning) or dropping it entirely.
In the end, Day Fifteen turned out to be too long to bundle with another post, and the idea of using data for good was (and is) important to me, so it stayed the way it was.
We set the bar kind of high with days eleven, twelve, and thirteen. I haven't had a lot of success with turning out breakthroughs at quite the same rate. Should I expect to? What should be more important to this project, coherence or consistency?
On Saturday, I actually ran into Rachel Thomas who, along with Jeremy Howard, produced the Fast AI course that I've been writing about right here! Wow! San Francisco is a super cool place.
The seven lessons I'm working through right now actually make up Part 1 of Fast AI. I asked her when Part 2 was coming out, because on the website it says:
Part 2 will be taught at the Data Institute from Feb 27, 2017, and will be available online around May 2017.
And it's now July. She told me that Part 2 was basically finished, but they were taking a little extra time to make sure the videos were perfect. It's good to know that I'm not the only person who takes a little longer on projects than they thought they would!
She also asked me where I was at in the process, and apparently there are other people that spend a long time getting through the first two lessons, but it's supposed to get easier oh god I hope it gets easier.
Anyway, we met at an event hosted by Delta Analytics, an organization that helps nonprofits around the world solve data problems for free. I'd first heard of Delta when I met one of the (founders? directors? both?) for coffee thanks to the now defunct Weave (RIP) but that's not really relevant information.
The team presented a few of the things they'd been working on throughout the previous year, and I think objectively the coolest project was the one where they helped a nonprofit called Rainforest Connection catch illegal loggers in rainforests by taking audio files captured by solar-powered recycled cell phones, converting them into spectrograms, and building a neural network to detect the frequency of a chainsaw with image recognition.
Oh and the recycled cell phone rigs are called guardians and this is what they look like:
You can help out by donating your old phone.
But personally, my favorite project was the one where they sent a team to Nairobi to teach a machine learning course, which they're going to release online for free.
Self-empowerment through online learning? Bringing state-of-the-art to the developing world? You had me at "underserved communities". I had the opportunity to share what I knew about online marketing with the startups at MaGIC a while back and it's still one of the most meaningful things I've ever done.
One thing that does still worry me is the continued consolidation of the world's data in the hands of a shrinking number of entities. Training a bunch of data scientists around the world will ultimately do very little to help their communities if they all end up working for the same handful of corporations headquartered in Silicon Valley and Shenzen.