The deep learning space has seen a tremendous level of innovation in the last couple of years. From new versions of established frameworks such as Caffe, Torch or Theano to new entrants in the market such as Microsoft Cognitive Toolkit or the ultra-popular TensorFlow, not a week goes by without some important development in the big learning space. The rapid evolution of deep learning frameworks has enable its adoption but some of the top artificial intelligence(AI) stacks in the market such as DeepMind which was originally built on Torch and recently switched to TensorFlow.
Despite the excitement about deep learning technologies, most of the relevant developments are taking place in the consumer or B2C markets while the enterprise remains cautious about the architecture and models that need to be put in place to adopt big learning technologies. Part of the slow progress in the enterprise is caused by the fact that some deep learning cloud platform s offer a very easy, entry point fro the implementation of basic deep learning scenarios such as image recognition or sentiment analysis which cause many organizations to initially leverage those cloud deep learning services instead of more complete open source deep learning frameworks. The second and most important factor contributing to the slow adoption of deep learning frameworks in business solutions is the fact that, currently, there are no robust enterprise distributions of those technology stacks.
Driving some lessons from popular open source technology movements in the enterprise such as Hadoop or Docker, we can easily see that the deep learning market could benefit from some enterprise platforms that expand frameworks such as TensorFlow, Caffe or Torch with enterprise-ready capabilities. From that perspective, the market needs the equivalent to a Cloudera or Hortonworks for the deep learning frameworks. The more interesting question to explore how would enterprise distributions of deep learning frameworks look like.
Five Capabilities of Enterprise Deep Learning Distributions
Copying some ideas from popular enterprise distributions of open source technologies as well as from the deep learning cloud platforms, I’ve put together a few ideas of capabilities that should be considered relevant in an enterprise deep learning distribution:
1 — Enterprise support: Different support models are key in order to make enterprise comfortable with the adoption of deep learning frameworks.
2 — Management & Training Tools: Deep learning models require constant maintenance and training. An enterprise deep learning distribution should provide a toolset that enables these capabilities in a way that can be used by devops and domain experts.
3 — Model Monitoring: Tracking the performance of deep learning models as well as other relevant operational analytics is another essential capability for the adoption of deep learning frameworks in the enterprise.\
4 — Deep Learning APIs: Taking a page out of the cloud deep learning platforms playbook, I believe an enterprise deep learning distribution should enable seamlessly expose models via APIs that can be easily consumed by third party applications.
5 — security Security is an omnipresent element in enterprise distributions of open source technologies. An enterprise deep learning distribution should include relevant security capabilities such as authentication, access control, auditing while also integrating with mainstream enterprise security infrastructures such as Microsoft Active Directory .
Who Will Bring Deep Learning Frameworks to the Enterprise?
From the existing players in the market, I believe the enterprise big data platforms such as Cloudera, Hortonworks or MapR are on an enviable position to deliver an enterprise distribution of open source deep learning frameworks. Of course, we can always count on new startups entering the space.