I'm negative on marketing fluff. You're free to enjoy it of course.
>> Umm just like how everyone else does it.
Wrong. Their ability to maintain and extend any of those libraries is limited. TensorFlow is hundreds of thousands of lines of code. Google can maintain and extend just fine. I doubt that H2O has the chops. That matters to their clients, because TensorFlow breaks things with new releases, and those clients will want someone to fix it. Do you call H2O? Do they patch Google into the conference line? I'm saying they're not in a position to offer commercial support. Anyone who has worked in enterprise will see the risks here.
>> What does your statement even means?
The algorithms that H2O specializes in do not benefit that much from the massively parallel computation that GPUs offer. If you don't understand the differences in the computation required by neural nets as opposed to random forests, you will not understand my critique. Neural nets offer superior accuracy on many problem sets and often need GPUs. Random forests do neither of those things.
H2O, for most of its long existence, has operated purely on CPUs just fine. But GPUs and deep learning are hot, so they are associating themselves with the buzzwords, even though they are not really in a position to a) build deep learning solutions or b) make their real product perform well on GPUs. Marketing fluff.
>> You mean libraries?
Why yes, I mean libraries. Thanks for clearing that up.
There's a lot of noise in the deep learning and AI space. Watson is one egregious example. H2O is another. While the technology itself is promising, some of the vendors are dubious.
In the end, most AI/DL implementations in products are fluff. It feels like a lot of products are about "HEY LOOK AT ME, I CAN DO IT TOO!", which screams immaturity of the concepts.
From personal experience, many customers ask about our capabilities - but when it actually comes to using any of them there are only a handful of actually relevant ML algorithms that are useful, and those used to be called 'statistics' up until fairly recently.
>> Umm just like how everyone else does it.
Wrong. Their ability to maintain and extend any of those libraries is limited. TensorFlow is hundreds of thousands of lines of code. Google can maintain and extend just fine. I doubt that H2O has the chops. That matters to their clients, because TensorFlow breaks things with new releases, and those clients will want someone to fix it. Do you call H2O? Do they patch Google into the conference line? I'm saying they're not in a position to offer commercial support. Anyone who has worked in enterprise will see the risks here.
>> What does your statement even means?
The algorithms that H2O specializes in do not benefit that much from the massively parallel computation that GPUs offer. If you don't understand the differences in the computation required by neural nets as opposed to random forests, you will not understand my critique. Neural nets offer superior accuracy on many problem sets and often need GPUs. Random forests do neither of those things.
H2O, for most of its long existence, has operated purely on CPUs just fine. But GPUs and deep learning are hot, so they are associating themselves with the buzzwords, even though they are not really in a position to a) build deep learning solutions or b) make their real product perform well on GPUs. Marketing fluff.
>> You mean libraries?
Why yes, I mean libraries. Thanks for clearing that up.
There's a lot of noise in the deep learning and AI space. Watson is one egregious example. H2O is another. While the technology itself is promising, some of the vendors are dubious.