IBM PowerAI – Taming AI Complexity via Collaborative Innovation

By Charles King, Pund-IT, Inc.  May 10, 2017

The growing interest and excitement around artificial intelligence (AI) is natural enough. For enterprises and other likely customers, AI offers benefits ranging from gaining insights into often opaque information resources to automating critical business processes. For involved vendors, AI represents potential new income streams, along with numerous opportunities to differentiate their own solutions from competitors’ offerings.

But in the rush to commercialize AI, some vendors have resorted to fudging or obscuring substantial challenges that need to be addressed in order to successfully deploy AI solutions. The problem is that such willful oversimplification does no one any good; willing customers may eventually feel hoodwinked and so-called knowledgeable vendors can come to look like fools.

The fact is that AI deployments are seldom simple or easy, and denying those points tends to result in more harm than good. That’s one of the reasons that IBM’s new PowerAI deep learning software solutions are so intriguing. The company is offering new tools that should lower the barriers to deploying AI, but in doing so it is also helping to clarify the investments and commitment required for AI efforts to succeed.

Why AI is hard

What is so difficult about artificial intelligence? Like many other enterprise IT initiatives, AI depends on the seamless integration of complex IT components, as well as successful interactions between sometimes disparate individuals and teams.

Let’s consider the first point. As an example, enterprise AI deployments typically include:

  • High performance computing (HPC) server and storage systems that require effective management and fine tuning to reduce machine learning and deep learning training times
  • Capacious data lakes and stores leveraging technologies, like Hadoop HDFS and NoSQL databases with assets that need to be curated, prepped and tagged to streamline extract, transform and load (ETL) functions
  • Distributed computing solutions leveraging cluster management and machine learning framework scaling tools, along with technologies like Apache Spark and MPI
  • Machine learning and deep learning libraries and frameworks utilizing TensorFlow, Caffe or SparkML, as well as the knowledge required to choose the most effective libraries and frameworks for specific projects
  • Application programming interfaces (APIs) for specific AI functions, including speech, vision, natural language processing and sentiment detection. These can include commercial offerings, like those available through IBM’s Watson cognitive platform, as well as custom APIs that enterprises create for their specific needs
  • Applications for specific industry use cases that are capable of accessing the data necessary for AI processes and outcomes

Add to that the fact that AI deployment teams and team members are anything but homogenous. In fact, it is not uncommon for the data scientists overseeing projects work with developer teams and data center staff who initially have minimal experience with AI processes.

The requirements for AI processes can also vary wildly depending on the use case. For example, what works for applications that depend on data collected with complex sensor arrays, like those used in self-driving cars and computer vision, have little in common with those used for credit risk analysis and real-time fraud detection in the finance industry.

In other words, since AI projects depend on both common IT practices and uncommon applications and use cases, success requires seamless interactions and meaningful collaborations between a wide variety of individuals and teams.

IBM’s new PowerAI solutions

What is IBM bringing to the table with its new PowerAI software? Four innovative, potentially valuable solutions for its Power Systems customers.

  1. DL Insight, a new software tool that monitors deep-learning training processes and recommends parameter adjustments to enhance training performance
  2. A version of the TensorFlow machine-learning framework (built by Google) optimized for Power Systems distributed environments leveraging NVIDIA’s Tesla GPUs and NVlink technologies
  3. Integration with IBM Spectrum Conductor for Apache Spark that speeds preparation of the data sets used in AI application training, and
  4. AI Vision, a new software tool that enables developers who have limited knowledge of deep learning to effectively train and deploy deep-learning models for adding computer vision functions to their applications

It should also be noted that these new offerings are merely the latest evidence of IBM’s longstanding support for and substantial investments in artificial intelligence and cognitive computing. In fact, if one considers the above list of inherent AI challenges, it is possible to detail significant progress and notable innovations IBM has delivered in each and every area.

Those include the company’s Watson technologies which easily qualify as the market’s best known and most fully enterprise-class cognitive platform. However, it is also unsurprising that these newest solutions found their genesis in the company’s Power Systems. Not only are those solutions the foundational hardware architecture supporting Watson but its robust performance, capacious memory and other enterprise-class characteristics make Power Systems ideal for supporting a wide range of compute-intensive workloads, including AI.

That is positive news for IBM, of course, but it is also likely to impact many of the 300+ members of the OpenPOWER consortium that are developing new and advanced data center solutions based on the company’s POWER silicon. In other words, these newest advances by IBM’s Power Systems organizations should trigger ripple effects that extend well beyond the company to numerous other AI projects and involved parties.

Final analysis

Overall, IBM’s new PowerAI deep-learning software offerings are an intriguing mix of new and enhanced technologies designed to address some fundamental challenges facing enterprises pursuing AI development. In particular, the new AI Vision and DL Insights tools seem likely to enable and enhance artificial intelligence efforts in organizations whose IT organizations may lack deep experience in AI development.

In other words, along with taking some notable technological steps forward, IBM is also addressing some of the inherent organizational and cultural issues that hinder some AI collaborations. Some may dismiss IBM’s solutions as simplistic but their inspiration comes from a company that has achieved remarkable success in AI development, tangibly understands the challenges and difficulties involved and is using its experience to simplify and ease others’ way forward.

© 2017 Pund-IT, Inc. All rights reserved.