Starting off the talk with some garage robot hacking from @l2k using @TensorFlow oreilly.com/learning/how-to-…
BayLearn 2018 talk videos (youtube playlist) youtube.com/watch?v=I9dALofe…
Key gotcha/trick: blog post experiments showing near-linear speedup increase batch sizes as number of workers increase (which may cause failure to converge). We need to see convergence plots! – at San Francisco, CA
Key idea: focus your precious training iteration cycles on more promising configurations. – at Financial District, San Francisco
3
2
Thanks again to speakers @cem3394 (@useTakt) & Cliff Szu (@hadron_cloud), host/drinks sponsor @useTakt, books/food sponsor @OReillyMedia, and of course everyone who came out!
4
What does @hadron_cloud do? Run in web browser to scavenge idle cycles for CNN training, then compensate user *with cryptocurrency*...! However: NOT just solving (worthless) inverse hash problems.
Cliff Szu (@hadron_cloud) dropping mobile adoption factoids: avg Indian data use already 10x of USA, 2025 smartphones > humans, ... , plus: increasing edge compute power on the device itself.
1
Videos posted for 11/21 meetup!
-"Scaling up ML @Square" by @CorinQi vimeo.com/244240576
-"@deltanalytics - data for social good" by @JJXWang vimeo.com/245062455
Lots of cool algorithmic work in @RainforestCx illegal logging detection and localization, eg Independent Component Analysis (ICA) to try isolating chainsaw sound vs other background noise.
Work by @RainforestCx and @deltanalytics on using simple audio sensors and supervised ML to detect illegal rainforest deforestation: see chainsaw vs birdsong spectrograms
The two key programs by which @deltanalytics leverages ML/data tech to help drive positive social outcomes: "data service grant" and "ML for good"
1
The 3 major dimensions of complexity for scaling ML applications @Square: wide variety of products/sellers, development/iteration cycles, and operational deployment.
1
Replying to @SFMachineLearn @Andrewdblevins














