To be honest, I'm not sure how Gneural plans to compete with those packages without support from CUDA or cuDNN, all of which are distinctly not open source.
The Linux kernel undoubtedly many features that the Hurd system lacks, but that is due to the severe lack of manpower of the latter system and the billions of Dollars being poured into the former.
On the other hand the Hurd has features that the Linux kernel can never hope to achieve because of its architecture.
That's why GNU Hurd is essentially a dead project. Sadly it never attracted the attention and manpower necessary for it to survive.
> On the other hand the Hurd has features that the Linux kernel can never hope to achieve because of its architecture.
Also think of the effort it took to introduce namespaces to all the Linux subsystems. After a decade the user namespace still has problems. This is ridiculously easy on a distributed system, yet very hard on a monolithic one.
I don't see the point either. Gneural will probably never be better than Theano, Torch, Tensorflow, Caffe, et al., which are already open. If anything, time/resources are much better invested in contributing to a polished/competitive OpenCL backend to one of these packages.
It seems like there must be more at play, but I'll admit a lack of insight and imagination on this one.
I think the reasons are twofold: 1. CUDA had a big headstart over OpenCL. 2. NVIDIA has invested a lot in great libraries for scientific computing. E.g. for neural nets, they have made a library of primitives on top of CUDA for neural nets (cuDNN), which has been adopted by all the major packages.
AMD should have invested much more heavily into ML, if they had, their share price would probably look a bit better than it does now.
This looks interesting - running CUDA on any GPU. http://venturebeat.com/2016/03/09/otoy-breakthrough-lets-gam...
Also somewhat related: AMD seems to be moving towards supporting CUDA on its GPUs in the future: http://www.amd.com/en-us/press-releases/Pages/boltzmann-init...
Its sort of supporting CUDA, just like a car ferry sort of lets your car 'drive' across a large body of water.
Also, the premise of OpenCL is somewhat faulty. You end up optimizing for particular architectures regardless.
Yeah, this is one reason I'm really hoping some of the stuff AMD is pushing, in regards to openness around GPUs, gains traction. And why I am hoping OpenCL continues to improve so that it can be a viable option. Being dependent on nVidia for all time would blow.
I don't think this is wrong, per se, but it is ...funny when the fsf portrays their work as morally superior to us horrible corporate permissive license lovers, while inexorably depending on non-free components.
In an ideal world this project will be popular and will lead to someone on gnueral writing nvidia compatible drivers that will allow them to reject nvidia's, but I'm not optimistic. Not because of some incompetency on the Gnueral team, but nvidia's long history of making life very difficult for open driver writers.
It will be prohibitively difficult to train the model without some kind of hardware assistance (CUDA). This means that if we're building an ImageNet object detector, even if the code implements the model correctly the first time, training it to have close-to-state-of-the-art accuracy will take several consecutive months of CPU time. Torch has rudimentary support for OpenCL, but it isn't there yet. There are very good pre-trained models that are licensed under academic-only licenses that also help fill the gap. (This is about as permissively as it could be licensed because the ImageNet training data itself is under an academic-only license anyway.)
I'm not sure what niche this project fills. If you want an open-source neural network, you have several high-quality choices. If you need good models, you can either use any of the state-of-the-art academic only ones, or you would have to collect some dataset completely by yourself.
Does this necessarily follow, that a machine-learning model is a derived work of all data it's trained on? As far as I know, the law in this area isn't really settled. And many companies are operating on the assumption that this isn't the case. It would lead to some absurd conclusions in some cases, for example if you trained a model to recognize company logos, you'd need permission of the logos' owners to distribute it.
(This is assuming traditional copyright law; under jurisdictions like the E.U. that recognize a separate "database right" it's another story.)
I'd like to note that some publishers, like Elsevier, allow you access to their dataset (full texts of articles) under a license with the condition that you can not freely distribute models learnt from their data.
But the CPU fallback is there
OTOH, it obviously matters a lot if you're constantly iterating and training multiple times a day or whatever.
It takes a week to train a standard AlexNet model on 1 GPU on ImageNet (and this is pretty far from state of the art).
It takes 4 GPUs 2 weeks to train a marginally-below state of the art image classifier on ImageNet (http://torch.ch/blog/2016/02/04/resnets.html) - the 101 layer deep residual network. This would be 20 weeks on an ensemble of CPUs. (State of the art is 152 layers; I don't have the numbers but I'd guess-timate 3-4 weeks to train on 4 GPUs).