DBTB INT Chris Fregly

my name is Chris fregley research scientist at flux capacitor AI here in San Francisco will start up I just being surrounded by you know it was actually hard speaking at the time I spoke at because there were two other sessions that I wanted to see so just being surrounded by good speakers and I get to sneak into the speaker lounge which is cool and talk to those guys and yeah people ask great questions this is one of the best grassroots comp riley conferences i yeah that I've been to I think my first one was three years ago in a small place over at fort mason i think it was and yeah just kept going from there and so I spoke last year and I spoke this year and it's been awesome about looking forward to it next year for me personally the reason I got into spark and the machine learning and data science and tensorflow is just the math I mean at the core of it you know we're all computer science where we're all right like math based and you know logical and things like that so it starts like activate a lot of things that the feelings I had in college and things like that and you know high school learning calculus learning physics and that kind of thing and it's more than just data pushing but it's actually an early data transformations it's more sort of deriving or like value out of these things and finding hidden features between interactions and you know finding clusters within data that you didn't see you know from rarely not until you actually run the numbers and yeah they run the algorithms so cool stuff specifically from my talk I cover quite a bit of ground so some of the things I focused personally are thinking beyond right like the exact values and thinking more about approximate and like approximations and especially at you know high scale high-throughput streaming rally depending on relig actual use case you can sort of really get away with right like approximations and probabilistic methods versus like getting the exact value so and sort of thinking you know yeah just beyond again like data pushing and let's see some other things I focus on are just being able to take a dataset and instead of just just piping into these algorithms but actually starting to write a kind of look at the data first and see if you can you know find structure there right before you actually just just start throwing these into the algorithms see what else my particular talk I dive deep into like a lot of sort of the spark sort of internals in there and right like tear apart project tungsten and like figure out what's going on with the Java Virtual Machine and rightly down to the sea level and rarely down to the operating system level so uh let's see I just talked to a buddy of mine from Apple about how he became a data scientist there is pretty interesting basically if you start as an analyst and you know you you just you just have to analyze data for a couple years and kind of get your feet wet and then you can start to expand and do some of the more advanced type stuff there you know I yeah going from really good to great I guess you just have to read and read and read everything until the things that you read start to reference where the other stuff that you've read right so you can keep going and going and going right but like once you sort of hit this or a stable state where you're like okay I have the intuition so there's first building the intuition you know which is kind of not getting bogged down with the actual details of the algorithm but just just sort of step back and you don't think about what is going on and why like you would use a particular method over some other method that's kind of the the first hub and then the great is just just being able to get there really quicker being able to not really rely on libraries that you've been given but have the confidence to actually go in and start to make tweaks for your specific domain rightly versus just the generalized like models given to you by spark and tensorflow and things but actually get in there and yeah like tweak them you