In a previous article, I discussed the emergence of cloud deep learning platforms such as Floyd and its viability on a market dominated by incumbents such as AWS, Google Cloud, Azure or IBM bluemix. Today, I would like to focus on the potential capabilities that these platforms should explore in order to be successful in the market.
From a functional standpoint, technologies such as Floyd offer a cloud service to automate the execution of deep learning from from popular frameworks such as Caffe, TensorFlow, Chainer and many others (over 10). Additionally, Floyd expects to build a marketplace for curated datasets and algorithms. While Floyd’s value proposition is certainly comprehensive, there are some serious challenges for that type of platforms on a market that, although young, is already dominated by incumbents such as Amazon, Google, Microsoft and IBM (see the previous article). However, technologies like Floyd also have an opportunity to capitalize on the well known limitations of the incumbent platforms. In order to do that, there are a series of capabilities that should be considered as part of their roadmap.
Capabilities of Standalone Deep Learning Cloud Platforms
1 — Support for Multiple Deep Learning Frameworks
While cloud cognitive APIs are a great potion for simple deep learning tasks such as image recognition, many artificial intelligence(AI) scenarios require custom deep learning models. For those scenarios, deep learning frameworks such as TensorFlow or Caffe offer incredibly robust capabilities. Providing a runtime and infrastructure for the execution and management of programs authored on different deep learning frameworks should be a key capability of a cloud deep learning platform. This feature can also result on a strong differentiator as most incumbent cloud deep learning platforms don’t offer support for those frameworks( with the exception of Google cloud ML that supports TensorFlow programs).
2 — Training and Management Tools
The current generation of open source deep learning frameworks is extremly sophisticated when comes to the implementation of AI programs but it lacks the operational toolset to operate in enterprise environments. Tools for training, monitoring and managing deep learning models should be an essential element of cloud deep learning platforms.
3 — Hybrid Infrastructure
To address a broader spectrum of enterprise scenarios, a deep learning platforms should be able to operate on both cloud and on-premise infrastructure. That level of symmetry is missing in the current generation of PaaS incumbents.
4 — Models as APIs & Data Marketplace
Providing a curated data source catalog will allow data scientists to author deep learning models without spending time and effort provisioning and curating the data. After the models are completed, they should be exposed via APIs that can be consumed by third party applications.
5 — Self-Service Data Science Tools
Environments such as Jupyter, Zepellin or Google dataLab have become extremely popular among data scientists. I believe that a cloud deep learning platform should leverage similar concepts and enable an interactive, self-service experience to author deep learning models using frameworks such as TensorFlow or Theano against curated datasets.
In the current AI market climate, we are experiencing an M&A frenzy. As a result, it is conceivable that cloud deep learning platforms will become hot acquisition targets. Surprisingly, I believe that PaaS incumbents such as Google, Microsoft or IBM are unlikely to pursue other cloud deep learning platforms as their technology stacks and visions don’t quite align. Having said that, there are several groups within the AI space that can become really acquisitive when comes to cloud deep learning platforms:
1 — Big Data Distributions: Big Data platform vendors such as Cloudera, Hortonworks or MapR are looking to expand their offerings beyond infrastructure and into higher level data services. Cloud deep learning platforms are an emerging market that naturally complements the capabilities of these platforms.
2 — Second Tier Cloud Providers: Cloud platforms such as Alibaba Cloud or Oracle Cloud need to quickly expand their AI offerings. Acquiring solid deep learning services could be a great way to become more competitive against the incumbent PaaS vendors.
3 — Amazon & Salesforce: AWS deep learning stacks are still training the competition and they can become a really compelling acquirer (many cloud deep learning platforms already run on AWS). Salesforce recently partnered with IBM Watson to develop its Einstein service. However, adding deep learning capabilities to Heroku is still a viable idea.