Deep Learning and Artificial Intelligence Platforms are Gaining Traction Behind the Firewall
Artificial intelligence(AI) and deep learning(DL) are two of the most important items in the agenda of modern enterprises. Until now, most of the complete experiences for building AI-DL solutions have been delivered as part of cloud services such as Watson Developer Cloud or Microsoft Cognitive Services but now the attention seems to be shifting to on-premise environments.
Don’t take me wrong. we have had plenty of incredible innovation in open source DL and AI frameworks that can be used on-premise. However, the on-premise infrastructures and tools required to scale and manage large AI-DL workloads are still limited compared to its cloud counterpart.
While AI-DL cloud services are a great starting point for relatively mainstream scenarios, most serious enterprise AI-DL use cases require custom algorithms and models that can only be implemented using AI-DL frameworks. Additionally, a large percentage of the data required on enterprise AI-DL scenarios resides behind corporate firewall. As a result, there is a strong demand for on-premise AI-DL platforms that can operate an an enterprise scale. While companies such as how.air have done a remarkable job enabling AI-DL applications, there is plenty of room and demand for new stacks.
Last week there were two announcements that indicate that we are about to see a lot more activity in the on-premise AI-DL platform space. First, Yahoo announced the open source release of TensorFlowOnSpark, a framework of executing TensorFlow programs as part of a Spark infrastructure. Also last week, IBM announced its intentions to bring Watson-like AI-DL capabilities to on-premise environments. Both announcements are clear indicators that on-premise AI-DL solutions are going to see more attention and innovation in the near future.
TensorFlowOnSpark combines two of the most popular platforms in the advanced analytics space on a single, cohesive experience. Using the new framework, TensorFlow applications can leverage Spark capabilities in areas such as advanced data computations, scalable infrastructures, SQL data access, data streaming, R supports and many others. Additionally, TensorFlowOnSpark allows TensorFlwo programs to be managed and monitored using standard Spark tools. The release of TensorFlowOnSpark follows CaffeOnSpark, another initiative by Yahoo to adapt FL frameworks to advanced data computation platforms such as Spark or Flink.
IBM’s announcements addresses one of the major requirements of Watson’s customers. Conceptually, the solution attempts to provide a symmetric set of capabilities between on-premise and cloud environments. The proposed solutions will work with mainstream tools such as Spark ML, TensorFlow or H2O as well as it would provide support for different programming languages such as Scala or Python. One of the most innovative additions to the stack seems to be IBM Research’s Cognitive Assist for Data Science which helps data scientists select the right algorithm based on specific requirements.
The announcements by IBM and Yahoo share the same pattern combining advanced DL frameworks with scalable data computation platforms such as Spark or Flink. That model could become the most viable short term option for adapting DL technologies to mission critical enterprise environments.