Machine Learning: A Full Stack View


According to Gartner’s Hype Cycle, machine learning is the hot new trend in the technology industry.  Why all the hype and excitement around artificial intelligence, big data, machine learning and deep learning?  As many of us in the industry know, machine learning and neural networks are certainly nothing new.  The buzz however this time around is being driven by the confluence of multiple factors including bigger data (and more importantly, labeled data), advances in scale compute, algorithmic innovation – and most importantly, killer apps that can take advantage of a data explosion.  In a world with billions (and in the near future, tens of billions) of connected devices, the amount of unstructured data that is collected by large organizations has quickly become unmanageable by traditional analytical techniques. Machine learning (ML), and its related branch, deep learning (DL), provide excellent approaches to structuring massive data sets to generate insights and enable monetization opportunities.

Generally speaking, machine learning is a set of algorithms that learn from data.  However, ML these days isn’t your father’s simple regression technique that might have worked well on smaller data sets.  The explosion of unstructured data requires new algorithms to process it, and the ML/DL ecosystem is evolving quickly in support.  From a Deep Learning perspective, a great example of this is the recent Microsoft Research Imagenet winner, the 152 layer Residual Net.  This massive neural network has an amazing amount of representational power and actually outperforms human level performance on many visual recognition tasks.

These types of algorithms actually perform better the more data is consumed making them a perfect match for the unending amount of data created today, assuming of course we can efficiently annotate it. From an application perspective, ML is not limited to ImageNet and object recognition.  It has already changed the way we shop at websites like Amazon and the way we are entertained by services like Netflix. ML is also being leveraged by cyber security applications to adapt quickly to threats and financial services institutions for highly accurate fraud or insider trading detection.

To quote Sundar Pichai at Google, “Machine learning is a core, transformative way by which we’re re-thinking how we’re doing everything.“

Because of this, Intel is investing heavily to enable the industry by providing a full stack solution for everything from highly scalable and efficient hardware to tools and libraries that will ease development and deployment of machine learning models into applications.

Starting at the lowest level, Intel is optimizing its hardware to target the highest single-node and cluster level performance including compute, memory, networking, and storage. This work builds on the capabilities of the Intel® Xeon® and Intel® Xeon Phi™ processor families, Intel® Solid-State Drives, new 3D XPoint memory technology, and Intel Omni-Path Architecture.  Our Intel Scalable System Framework (Intel SSF) configurations are designed to balance these technologies and efficiently and reliably scale to increasingly larger data sets.

Moving up the next level of the stack, a set of highly tuned and optimized libraries are required to truly extract maximum performance out of the hardware.  Enhancements and additions are being made to the Intel Math Kernel Library, which provides a set of tuned math primitives, and the Intel Data Analytics Acceleration Library, which optimizes and distributes a broad set of machine learning algorithms.  These libraries also abstract the complexity of the underlying hardware and instruction set architecture (ISA) providing a level of programming that is comfortable for most developers while still highly performant. In addition to enhancing the libraries themselves, we are actively integrating with and contributing code back to key open source projects that are influential in machine learning.  This includes seminal projects like Caffe from UC-Berkeley, the Apache-Spark project, Theano from the University of Montreal, Torch7 which is used by Facebook and Twitter and others like Microsoft’s CNTK and Google’s Tensor Flow.

On an even broader front, Intel is accelerating enterprises and application developers looking to use Machine Learning through the open source Trusted Analytics Platform (TAP) project which provides everything from big data infrastructure and cluster management tools to model development and training and application development and deployment resources. To further reduce friction for developers, TAP works with or is pre-integrated with popular frameworks and toolkits such as Spark-MLLib, H20, DL4J from Skymind and DataRobot to name a few.

For a deeper dive into Intel’s strategy, libraries and recent customer activities in the machine learning space, you can explore the slides from a machine learning session at the recent 2016 Intel Developer Forum in Shenzhen China. You can also access the video from a talk I gave in late March at the 2016 Hadoop + Strata World in San Jose.

Please stay tuned for more announcements and initiatives throughout 2016 from Intel regarding machine learning!