This is very exciting to see these distributed deep learning frameworks open sourced by top companies. What I do not understand is why amazon AWS and other top cloud service are not integrating these frameworks in their services. Training a distributed neural net should be as simple as defining the model and specifying the resources.
agreed - though I'm somewhat saddened to see that this library doesn't seem to be getting the same positive reception (in terms of scoring / comments) as Tensorflow did yesterday. Am I right and if so, is there a technical reason for that?
One technical reason may be that C++ (I only looked at LightLDA) has a higher learning curve than Python, which the Tensorflow docs stress.
Though Tensorflow stayed at #1 for a while, it only garnered around 200 comments, which is very high for a ML topic on HN (in my anecdotal experience, ML topics are highly rated but under-commented.) I imagine the audience for this library to be very small compared to Tensorflow, which likely included more ML/Google FOSS enthusiasts than day-to-day practitioners. Looks like this library is firmly targeted toward the latter.
Not sure if this was ready for public release.. The chess demo released this morning seems more polished than this library.. Also, probably not the best decision to post it a day after the TensorFlow release..
You do realize there are a fair amount of machine learning algorithms that do not run efficiently on the GPU right? Deep learning isn't the only method out there...
distributed ml tends not to work well on aws. Many (most?) distributed ml algorithms are essentially limited by network speed. In a p2p topology, many are effectively limited by the slowest peer-to-peer link of the (nodes choose 2) links. aws networking is mediocre at best; even the rack-local instances are relatively slow.
cred: helped build a distributed in-memory ml toolkit
This is very interesting. I was very surprised that Tensorflow did not use a central parameter server, and this looks like a good foundational parameter server to build something like DistBelief on top of from Microsoft. I wonder if they have a Deep Learning module that they will open source eventually. Microsoft has a few tricks they used in some of their recent papers and it would be interesting to see them in a production quality system.
I was doing some research for a small demo. As a proof of concept, implementing a small ConvNet using WebGL shaders. But in the course of my search I stumbled on this very broad and interesting patent granted to Microsoft Corp.
Processing machine learning techniques using a graphics processing unit
Pretty amazing when you consider they wrote the application a decade ago! There is no arguing we are experiencing a golden age and an embarrassment of riches provided by ML toolkits and GPU cloud capabilities. But further inspection of patents in the space brought up this gem, in which Emotient is attempting to patent the crowdsourcing of training data...
Collection of machine learning training data for expression recognition
That they're providing a distributed LDA implementation with and O(1) Gibbs sampler is a big deal. I haven't played with it yet but the numbers they are reporting relative to cluster size is orders of magnitude improvement.
