3 ms·
Just to give a different perspective. We are an on prem shop that sells to banks, telcos etc who can't use the cloud for compliance reasons. We make it work by
by agibsonccc 10y ago
Just to give a different perspective. We are an on prem shop that sells to banks, telcos etc who can't use the cloud for compliance reasons. We make it work by doing other things people need in those environments well. One example is my willingness to do an install via dvds rather than "the cloud".
We mainly do fraud detection and security related work.
We have also seen operations workloads (forecasting when machines are going to break or preventative maintenance)
Most of our core business isn't even CNNs.
One thing that's missing from the narrative is that researchers who vision and speech because it's "pretty" and you can demonstrate results on a large feature vector. It's also very relatable for normal people.
Most of corporate america (not silicon valley) has more traditional things like time series data. Not images.
I would say here that deep learning use cases aren't explored by most people and that there are other areas besides what the marketing with self driving cars is perpetuating.
The other thing I would add here is when companies get bigger they typically need to take core competencies like vision and speech in house. It can be hard to justify outsourcing your core business to google as you grow.
This trend however might change over time. I would love to hear contradictory statements here.
Disclosure: I am a deep learning founder in the same batch as the authors of this blog post.
- zzleeper 10y agoIf you were to use deep learning in-house for smallish problems, would you still use tensorflow plus a batch of GPUs, or something different? I can't take stuff to the cloud for regulatory reasons, and in order to contract w/a startup like yours, I would first need to make a case for it (with useful examples that just need to be improved)
- agibsonccc 10y agoI am the wrong person to ask. We compete with tensorflow going after the larger scale stuff where hadoop is a requirement and the cloud is a swear word. I work with .net and java shops where running on windows is a requirement and hippaa is considered "lite" and they dont know what a gpu is. Despite the marketing noise that is still most of the world. Look no further than another comment that showed java job postings vs machine learning. My biased comment out of the way: if you can convince ops on why you need this deep learning thing then be my guest. Their first question to you is probably going to be: Does dell/hp sell this as a reference box? It mostly comes down to roi and what your existing stack is. My job is hard enough as it is. If you can convince IT to allocate an r&d budget be my guest. That is more or less our specialty. Email is in profile if you want specifics.
- robotresearcher 10y agoYes, that might work fine, depending on the scale of your problem. Grad students do interesting work with small setups, and companies can too.
- plinkplonk 10y ago"One thing that's missing from the narrative is that researchers who [<--verb missing here-->] vision and speech because it's "pretty" and you can demonstrate results on a large feature vector."
- agibsonccc 10y agoSorry! Was writing this quickly. Good find. "do research using vision and speech" Too late to edit.