4 ms·
No. Early google had several machine learning systems including Phil, RePhil, SERI, SmartAss (which made google many billions before 2010), and Sibyl all of wh
by dekhn 5y ago
No. Early google had several machine learning systems including Phil, RePhil, SERI, SmartAss (which made google many billions before 2010), and Sibyl all of which predate 2010.
Rephil: https://www.quora.com/How-does-Googles-topic-model-Rephil-work https://www.quora.com/How-does-Googles-topic-model-Rephil-wo... and 26.5.4 Case study: Google’s Rephil 928 from Kevin Murphy's ML book.
Rephil was critical for google when they needed to cluster topics and it was used for a wide range of other things.
Smartass: http://static.googleusercontent.com/media/research.google.com/en/us/pubs/archive/41159.pdf http://static.googleusercontent.com/media/research.google.co...
Smartass is really weird if you come from the "batch training of deep networks" world, because it's online, and visits the dataset once, in temporal order. But. it's as ML as it gets; it was just turning joined clicks (between search queries and ad impressions) into weights for pCTR.
Sibyl was Google's secret sauce before tensorflow started in 2007, it implemented parallel boosting using mapreduce as a worker engine to do linear and logistic regression and was used to train the models for youtube watch next and play store (both of which produced enormous revenue at a time when people had no idea that youtube and android were gold mines) but predated those uses by at least 5 years. It was one of the oddest systems I've ever worked on and you can see some of the ideas it innovated in TFX today.
There's also SETI which IIRC even predated Sibyl, and you still can see vestiges of its functionality when doing archeological work on the google codebase.
I worked on systems related to smartass, worked on sibyl, and later tensorflow/TPU hardware/software. I met a number of long-time engineers and asked them deep questions about ML use at Google predating tensorflow and gained a deep understanding (and even then, I've left out many details about systems I'm not aware of). Before tensorflow there wasn't much use of GPUs or TPUs, for several reasons but once Vincent Vanhouke stuffed a few GPUs in his workstation and showed fast training to Jeff Dean, that changed quickly.