Other stuff: (conference site, with videos)
MSR spends as much as the NSF in support of computer science-related research!
Bill Gates referred to "polymathic individuals who can understand things across multiple boundaries" as important drivers of innovation ;-)
Speech recognition had a big breakthrough in the last few years due to deep neural nets ("beyond shallow HMMs ..."). See this video.
Exponential growth in computing power (Moore's Law, considered more generally than just counting transistors) is in trouble, as predicted from simple considerations of quantum effects and feature size. Even the multi core workaround is running out of steam ("Dennard scaling").
Cancer genomics hub at UCSD supercomputer center -- storage cost $100/y/genome at 50k genomes (currently at 10k?). 3GB/s outbound. (Cancer genomics slides: Haussler, UCSC.) Data costs > sequencing costs. David Patterson: Using Big D to fight the Big C. "Biologists are mean. CS people are nice - we share data and code... A bunch of sociological issues we have to fix here." ;-) FaST-LMM -- epistasis GWAS w/Wellcome Trust data, 60B pairs of SNPs (Heckerman).
Andrew Ng: unsupervised deep learning (video), tutorial. $100k GPU farm beats 16,000 CPU Google project that independently evolved a "cat neuron" :-)
Mac to Windows ratio among participants' laptops roughly 10 to 1. (Ratio moderated a bit by MSR employees ;-)
Bill Gates Q&A:
2 comments:
Moore's Law (which is really about transistor density, not computing power) became defunct several years ago, maybe a decade. If it had been in effect during that time, your laptop computer would be operating at 20 to 30 GHz instead of 2 to 3.
I emailed Professor Ng about it and he pointed me to Adam Coates' ICML brief on it. It makes me chuckle, because economically speaking, it is probably $100k infiniband network with the GPU hosts on the ends. The cost of the networking gear probably dwarfs everything else.
Post a Comment