Machine learning - STH should get involved
- Thread starter Biren78
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
Are you thinking about Caffe + GPU compute? I actually discussed this when I was at Mellanox not to long ago.
This is very close to my work area as well. I work with Hadoop on IBM, Cloudera and MS and have done quite a bit too, at scale too.
If I can find some data that isn't classifed up to here *holds up hand* then later on in the year when workloads cool down, I can contribute as well.
If I can find some data that isn't classifed up to here *holds up hand* then later on in the year when workloads cool down, I can contribute as well.
I am normally a machine learning skeptic. Right now everybody is jumping on it big time since quantum computing could make it much more efficient at the snip of a finger very soon. Even a non-universal quantum computer like the D-wave.
I don't think anybody means to say that they think ML has more future generally than other software styles. The looming breakthroughs make everybody go wild so that they are ready when something happens soon. They can always go the other stuff later.
I am a bit torn. I still see the same problems in ML that it always had. On the other hand it would be an easy ride on the currently ongoing wave.
Let me ask this way: any conferences in Hawaii, Malta or Nijmegen this year?
I don't think anybody means to say that they think ML has more future generally than other software styles. The looming breakthroughs make everybody go wild so that they are ready when something happens soon. They can always go the other stuff later.
I am a bit torn. I still see the same problems in ML that it always had. On the other hand it would be an easy ride on the currently ongoing wave.
Let me ask this way: any conferences in Hawaii, Malta or Nijmegen this year?
Didn't even notice the date
... Sorry for necro.The quantum for me is still like fusion nuclear power. I will believe it works when it works. Imho, not till 2030+
They've also had the whole DNA-computers going for 15+ years and its also about 20+ years away still. Quantum could be closer though.
I don't wanna push M-L more than it can handle either, but I do think it finally has come to be accepted and more importantly, can be used and sold as projects in real business scenarios;
Background: A lot of what I do is data staging, ingestion, governance, architecture, data lake work (new for 2016 though), statistics and prep for visualization tools (Tableau mostly). But I also do quite a bit of R, some BigR (cause it cough does not work all that well cough) and some Spark and Python. I will be doing Python on Cluster on Cloudera very soon though as we just made a partner-isch deal with Cloudera.
Business wise, as a consultant, I can say that we have some trouble with smaller companies and public entities, but it improves every quarter with beliefs and case stories where M-L comes in and makes it through a full iterative business cycle and pushes decisions from reactive, to preventive to driven by it. I have a few things where it cannot be used unsupervised (as in, not automated) such as where human lives are at stake, but apart from that there are quite a few business cases where its pure gold, sales wise atm. Fraud and other anomality places.
I agree that M-L might not be all THAT important for a benchmark, especially not on GPUs yet, as it isn't really that usual "big hog" on performance; That is only when all the other 95% of the work has been set up and works as intended and all governance is in place etc etc.
And then by far the most I've seen, run their stuff on CPU and in-memory servers. GPU is mostly for HPC (Supercomputers) and/or research projects that can go on Amazon clouds because of public datasets etc. Have not seen on-premise GPU clusters for private/public sector computing yet.
The conferences reference went over my head. Is it a specific annual conference or some "hot and cozy" place for a conference?
Just some favorite places. If the conferences are in the right place I can get enthusiastic about anything
Yeah. Malta does not have mosquitoes, BTW. No standing fresh water anywhere. Good for outdoors "technical" discussions until 2 a.m.Ahhh right! ... Yup, I remember an IBM conference I suddenly had a very vested business case interest in too. It was in Miami in February and I live in northern EU where the sun is partially gone from October to March
I am heading to the STH Sunnyvale colocation this morning. One of the big search engines has multiple racks full of NVIDIA GPUs just down the way as a "lab". Dataset I believe is being held on a Pure Storage array.
While not "on-premise" it is certainly in an offsite colocation for them.
Nice! I wish I could live in the US but it won't happen for now. Something tells me that the bigger companies over there are more ahead in embracing/adapting new stuff like actual GPU accelerated stuff because they have business cases that calls for this real need of right-here-right-now answers in this caliber.
Patrick, have you had any experience with the Ryft One cluster nodes that run FPGA's next to standard stuff to gain ASIC-like speedups on certain Spark projects?
I saw that. I also bought:
Python Machine Learning: Sebastian Raschka: 9781783555130: Amazon.com: Books and
http://www.amazon.com/Data-Science-Scratch-Principles-Python-ebook/dp/B00W4DTP2A/ for plane rides this weekend.
Python Machine Learning: Sebastian Raschka: 9781783555130: Amazon.com: Books and
http://www.amazon.com/Data-Science-Scratch-Principles-Python-ebook/dp/B00W4DTP2A/ for plane rides this weekend.