The AI spring will herald a surge in machine-learning-related M&A
Machine learning is hot, and interest in it is widespread, with everything from autonomous vehicles to automated medical diagnoses being driven by it. This is due to a number of factors. First, compute power is available as never before, especially via the hyper-scale cloud providers. Second, data is available in unprecedented volumes, which means the algorithms have more to get their teeth into (machine learning lives or dies depending on the quantity and quality of the data available).
Third, organizations are coming to the realization that machine learning is the only way to drive efficiencies through automation because there is simply too much data for humans to manage and analyze on their own; they need the force-multiplier effect that machine learning can bring. All these factors are already driving M&A, and will continue to do so for a long time to come. This report looks at the M&A that has happened, and some of the areas where we expect more.
The 451 Take
We believe the artificial intelligence (AI) winter spanning 30 years or more is finally over, and we're in the AI spring. From autonomous vehicles to cancer diagnosis, developers are building applications to take advantage of machine learning and help humans do things more efficiently than we can unaided. Anywhere there is a critical problem that involves analyzing data at scales beyond that humans can handle using current analytics tools is ripe for being transformed by machine learning. It's rare for a technology to be attractive to the entire technology stack from silicon to applications, but that's how broadly applicable machine learning is.
Let's define what we mean by AI, machine learning and deep learning, which are the most often used phrases to describe what is going on, and are often used interchangeably – not always correctly. AI is the quest to build software running on machines that can 'think' and act like humans – what can be thought of as general AI. Machine learning is a subset of artificial intelligence focused on using algorithms that learn and improve without being explicitly programmed to do so. The algorithms take data as an input, and the output is predicted data or actions. The algorithms improve as they are exposed to more data. Think of this as narrow AI.
Deep learning is a branch of machine learning based on a specific set of algorithms that attempt to mimic the working of the brain in the form of multilayered neural networks. Deep learning enables machine-learning algorithms to learn features from an input of raw data. In other words, it not only learns how to use the features (for example, the edge of a person's face), but it also learns how to represent the age of a person's face from raw data.
Google's AlphaGo software is a good example of deep learning. It learned how to play the board game Go from scratch. It was never told any strategies or tactics; it just played the game thousands of times before it got very good at it, and thus in effect learned the tactics of the game. AlphaGo was written by Google's DeepMind team – the result of the January 2014 acquisition of the London-based company of the same name for about $400m.
The 451 M&A Knowledgebase reveals some trends in machine-learning deals. Three use cases stand out in terms of significant deals, but there are many others. They are image recognition, big data analytics and security analytics. The buyers vary quite a lot, from chip companies like Intel to software firms such as Salesforce to automobile companies such as Ford. This reflects the broad appeal of machine learning. In fact, we believe we've reached a point where any predictions about the future of software that don't include extensive use of machine learning will be wide of the mark.
Notable machine-learning deals
Intel's two most recent deals – that of Movidius in September and Nervana Systems in August – were both driven by machine learning. Movidius' low-power chips enable image recognition in use cases including automated navigation, physical anomaly detection and augmented reality. Such small, low-power embedded devices across the Internet of Things (IoT) benefit from sensing the world around them and acting in real time.
Nervana had built an engine specifically for accelerating deep-learning algorithms. Deep learning lends itself to being optimized in silicon because it is a series of parallel operations across the layers of the neural networks, so we can expect other chip deals to optimize deep learning. Intel is also working on the development of Knights Mill, a next-generation version of its Xeon Phi many-core family, specifically optimized for deep learning, and due to emerge in 2017.
Nvidia is the biggest chip vendor really focused on machine learning, and already has cloud presence in every major cloud out there, including Amazon Web Services, Google Cloud Platform, IBM Cloud and Microsoft Azure, with all those providers offering Nvidia's graphics processing unit (GPUs) chips as part of their compute engines. Nvidia's share price has perhaps reflected its status as the main chip play where machine learning is concerned, up 229% year to date.
Softbank's newly acquired ARM Holdings might well be a buyer, although it paid $350m for Apical in May 2016. AMD, CEVA and Imagination Technologies also have their eyes on this emerging marketplace and could be potential buyers. Targets could include Wave Computing, which is about to launch what it says are faster than the general-purpose GPUs build by the likes of Nvidia and yet more energy efficient than field-programmable gate arrays (FPGAs), which are seen to be more adaptable.
Graphcore, a UK-based startup that recently announced a $30m series A round with some illustrious backers, plans to bring what it calls Intelligent Processing Units (IPUs) to market in 2017. Mobileye is focused on autonomous vehicle assistance with its EyeQ proprietary chip. Qualcomm's presence in mobile with its Snapdragon processor and its FastCV software development kit offers some machine vision function. Vuforia, an augmented reality platform, was sold by Qualcomm to PTC in 2015, bringing an application development layer to augmented reality applications in its IoT offerings.
Salesforce has made a lot of acquisitions in the past two years – 16 in all – and at least half of those have been driven by machine learning. So it was no surprise that it made Project Einstein – its machine-learning layer within all its applications – the centerpiece announcement of this year's Dreamforce event earlier this month. Plus, it promoted the former CEO of MetaMind (acquired in April 2016) to chief scientist, and put him in charge of R&D.
Whether Salesforce has all the machine-learning smarts it needs for now, we'll have to see. But there are certainly plenty of other business application vendors out there that could use some, notably Oracle. Although it said it had infused its applications with machine learning in what it called Adaptive Intelligent Applications the same day that Salesforce announced Einstein, the announcement was a bit light on details in terms of what type of machine learning was being used and how.
Infor bought Predictix in June 2016 for its retail predictive analytics, but we suspect there are more areas of its application portfolio that could be enhanced with machine learning.
Security analytics is another form of data analytics; trawling through vast swatches of unstructured and structured data looking for some sort of anomaly or indication of malfeasance.
Elastic's purchase of Prelert in September, IBM's acquisition of IRIS Analytics in January, Fair Isaac buying QuadMetrics in June and Splunk's pickup of Caspida in July 2015 are all examples of machine-learning-driven security analytics of various types. Many other targets remain, particularly in the user and entity behavior analytics category, including Exabeam, Fortscale, Niara and RedOwl Analytics.
In addition to the targets we've already mentioned, there are myriad small companies out there doing something related to machine learning, which makes pinning them down hard. Some will be attractive for their platform, some as a talent acquisition.
Probably the biggest pool is in the area of advanced analytics. Among the most interesting players here are Ayasdi, BigML, Nutonian, Skytree and Yottamine Analytics. As is always the case in complex software, companies are working to make it simpler, and thus broaden its appeal – not everyone is comfortable coding models in R or Python. Startups DataRobot and DMWay are two that are trying to make that real for those wanting to deploy machine-learning-powered predictive analytics without having that knowledge.
Other vendors are focusing on data scientists and developers in other ways, including Deepsense.io, which has a way to manage and monitor data science, and visual Spark framework Minds.ai, which offers software to build and train neural networks more quickly. Nara Logics builds a synaptic network of explicit and inferred connections across data so it can be used by applications.
There are also some stand-alone companies still offering text analysis powered by machine learning, including CognitiveScale, Expert System, Lexalytics and Linguamatic, although most of these companies have also built applications on top of their platforms to solve business problems such as sentiment analysis in social media, drug discovery and customer engagement. Of those listed, only CognitiveScale is a recent addition; the rest have been around for 5-10 years.
Buyers could come from literally anywhere within the tech industry, and beyond what we previously thought of as the technology industry. Automobile manufacturers are still likely to need image recognition or other similar software. Large application vendors will be buying predictive analytics and other forms of advanced analytics – we expect Oracle to be active here, but also potentially SAP. Industrial companies will also get in on the act – witness GE's recent acquisition of machine-learning startup Wise.io. Infrastructure software vendors will be interested because machine learning has applications up and down the software stack. And chip companies that haven't already gotten involved must be scrambling to acquire machine-learning smarts.