Thursday, December 18, 2014

Weaver: a new dynamic graph framework

My friend Yaron Weinsberg sent me a link to Weaver, a new dynamic graph framework from Cornell.

Wednesday, December 17, 2014

Minerva: open source deep learning on GPU software from MS

I got this from my colleague Chris DeBois - Minerva is an open source deep learning software. If you ask what is new as there are tons of similar packages - Minerva claims to be able to support multiple GPUs on the same machine.

Here is the paper which describes Minerva.

Tuesday, December 16, 2014

Bitcoin conference: Jan 27, SF

My friend and colleague Ben Lorica, chief scientist of O'Reilly Media never rests.. Now how organize an interesting BitCoin related one day conference on Jan 27 in SF.

Readers of this blogs are welcome to use discount code USGR20.

Intel Xeon-Phi is trying to catch up on NVIDIA

I got this article from my colleague Matt from Walmart:
http://hips.seas.harvard.edu/content/micmat-python-scientific-computing-intel-xeon-phi
it seems that Intel is trying to catch up on NVIDIA by creating tools for GPU processing.

NVIDIA has some opinion on Intel's effort which is listed here.


Friday, December 12, 2014

Hardcore data science

My friend and colleague Ben Lorica, just sent me a link to the hardcore data science track is his organizing at Strata San Jose. Super interesting topics - in addition, you are welcome to use discount code GRAPHLAB20 when registering.

Tuesday, December 9, 2014

GraphLab's deep learning - the power of graph applied to images

GraphLab's deep learning - the power of graph applied to images

A couple of months ago we have released a deep learning toolkit for GraphLab Create. We just got a code contribution from Marian Moldovan & Enrique Otero, from Beeva.comwhich utilizes GraphLab deep learning toolkit in a new and exciting ways.

Marian & Enrique created a super awesome application. Imagine you have a repository of images and you would like to understand the relation between the images. The images they used are of buildings in Barcelona, as this work was created at the hacknight of papis.io, a predictive API conference in Barcelona.

Here is the first building:
Inline image 1


And here is the second building:
Inline image 2


What is the architecture transition that can explain this path? Using GraphLab Create it is easy to compute!

In a nutshell, they first extracted images features using the deep learning toolkit. Then they used a nearest neighbor model to create a graph of all the similar buildings:

Next, they used the graph model to find the shortest path between two interesting buildings (number 16 and 23)
The Ipython notebook to reproduce this example is available here.




Friday, December 5, 2014

Deep learning startup raises 8M$ round A

Fresh news from my colleague Chris Debois. Richard Socher's new startup MetaMind just announced 8M$ for forming a deep learning based solution company. So far they have two functionalities: identifying objects in images and classifying text.

Readers of my blog may recall that a year ago I have already written about etcml, the project which is now the basis of this new startup.

A related company is Alchemy which provides an API for performing similar tasks.

Other related companies are Superfish, which finds similar photos of pets, decor and clothing, as well as Cortica who identifies objects in images. The difference is that MetaMind provides tools for developers and not an end application.

By the way, in GraphLab we also provide deep learning functionality.