If there is one true indicator to measure the disruptiveness of a new technology, it’s certainly the public outpouring of fear and suspicion. If we use societal angst as a measure, the current renaissance of artificial intelligence (AI) is a good candidate for groundbreaking technological disruption. AI will change life as we know it, as Elon Musk, Bill Gates, Stephen Hawking and other great minds have told us. The widespread anxiety about the harmful consequences of AI applications is not an unparalleled reaction to technological change but rather an expression of the societal unease that commonly precedes the changes associated with new technologies and the vast potential that comes with them.
We’re looking beyond today’s IoT, toward a future where smart connected devices not only talk with each other but where they use AI to interact with each other on our behalf. This new global fabric of artificially intelligent things one day will be known as the AIoT, the Artificial Intelligence of Things.
As a discipline of mathematics – and, to a certain extent, of philosophy – AI lived in the shadows for more than six decades before public interest suddenly soared in the present period. One reason for the current publicity is that, for a long time, considerations on AI applications were purely theoretical, or at least science fiction. For artificial intelligence use cases to become real in the present IoT environment, three conditions had to be fulfilled:
It is apparent that the two latter requirements depend on each other, and that the breakthroughs in deep neural nets could not have occurred without a significant increase in processing power. As for the input: large data sets of every quality – vision, audio, and environmental data – are being generated by an increasing number of embedded IoT devices. Today, the flow of data grows exponentially. In fact, annual data generation is expected to reach 44 zettabytes (one zettabyte is 1 billion terabytes) by 2020, which translates to a compound annual growth rate (CAGR) of 141 % over five years. Just five years after that, it could reach 180 zettabytes.
Starting around 2015, when multicore application processors and graphics processing units (GPUs) became widely available, we’ve also commanded the tools to cope with these amounts of data. Parallel processing became a much faster, cheaper, and more powerful business. Add fast, abundant storage and more powerful algorithms to sort and structure that data, and suddenly there is an environment in which AI can prosper and thrive.
In 2018, neural network-trained AI voice recognition software is an integral part of a variety of consumer and industrial applications. Computing power is increasing by roughly a factor of 10 each year, mainly driven by new classes of custom hardware and processor architecture. This computing boom is a key component in AI progress and helping to make AI mainstream in the future. And this is just the beginning.
By classic definition, artificial intelligence is a rather unspectacular affair. In his groundbreaking 1976 paper Artificial Intelligence: A Personal View, British neuroscientist and AI pioneer David Marr states: The goal of AI is to identify and to solve useful information processing problems and to give an abstract account of how to solve it, which is called a method.
The small but decisive detail with AI is that the information processing problems it deals with have their roots in aspects of biological information processing. In other words: AI looks to re-engineer the structure and function of the human brain to enable machines to solve problems the way humans would – just better.
Compared with the scientific effort, today’s industry’s approach on AI is much more pragmatic. Rather than trying to achieve a replica of the human mind, current AI development uses human reasoning as a guide to provide better services or create better products. But how does that work? Let’s have a look at the current approaches.
As a subset of artificial intelligence, machine learning uses statistical techniques to give computers the ability to learn without being explicitly programmed. In its most crude approach, machine learning uses algorithms to analyze data, and then makes a prediction based on its interpretation. To achieve this, machine learning applies pattern recognition and computational learning theory, including probabilistic techniques, Kernel methods and Bayesian probabilities, which have grown from a specialist niche to become mainstream in current ML approaches. ML algorithms operate by building a model from an example training set of input in order to make data-driven predictions expressed as outputs.
Computer vision is the most active and popular application field where ML is applied. It is the extraction of high-dimensional data from the real world to produce numerical or symbolic information – ultimately in the forms of decisions. However, until very recently, an extensive portion of hand coding was involved for machines to develop advanced pattern recognition skills. Human operators had to extract edges to define where an object begins and where it ends, apply noise removing filters or add geometrical information, e.g., on the depth of a given object. It turned out that even with advanced machine learning training software, it’s not a trivial task for a machine to make real sense of a digital reproduction of its environment. This is where deep learning comes into play.
It’s a decades’ old idea that software can simulate a biological neocortex’s array of neurons in an artificial “neural network”. Deep learning algorithms attempt exactly that – to mimic the multilayer structure and functionality of the human neuronal network. In a real sense, a deep learning algorithm learns to recognize patterns in digital representations of sounds, images, and other data. But how?
With the current improvements in algorithms and increasing processing capacity, we can now model more layers of virtual neurons than ever before, and thus run models in much greater depth and complexity. Today, Bayesian deep learning is used in multilayer neural nets to tackle complex learning problems.
However, what we can do today still mostly falls into the concept of “narrow” or “weak” AI – technologies able to perform specific tasks as well as, or better than, humans can. For instance, AI technologies for image classification or face recognition perform certain facets of human intelligence, yet not the full spectrum, or even a combination of several human capabilities. A machine capable of performing a multitude of complex tasks, one that exhibits behavior at least as skillful and flexible as a human being, would be considered “strong AI”. While the experts are divided over the question whether strong AI can ever be achieved, it doesn’t stop them from trying.
One solid indicator for the disruptive potential of the technology is the dimension of investment. According to McKinsey, $26 to 39 billion were invested in AI in 2016, most of it by tech giants such as Google™ and Baidu™. Being one of the most well-funded categories, Startups in the space increased about 141 %. Recognizing the industry’s enormous potential, governments around the globe aim to put themselves in the top positions of AI. While a notable amount of national AI plans are in creation, the greatest economic gains are expected for China (26 % boost to GDP in 2030) and North America (14.5 % boost), accounting for almost 70 % of the global impact.
According to recent studies, Artificial Intelligence has the potential to almost double the value of the digital economy to US$ 23 trillion by 2025. From a strategic viewpoint, AI’s biggest potential is seen in its complementary nature to the IoT. An integrated technology portfolio creates a powerful new platform for digital business value.
As we have seen, for AI to unfold its massive potential relies heavily on adequate hardware. Machine learning, in particular, requires enormous processing and storage capacity. A training cycle for one of Baidu’s speech recognition models for instance, requires not only four terabytes of training data, but also 20 exaflops of compute — that’s equivalent to 20,000 quadrillion math operations per second — across the entire training cycle. Given its hunger for powerful hardware, it is no wonder that AI today is still mostly confined to data centers.
Uncoupling AI from the data centers and advancing it to the endpoints of the IoT will allow us to tap its full potential. Edge processing has taken the control of computing applications, data, and services away from some central nodes (the “core”) to the periphery of the Internet. Processing this data at the edge, significantly decreases data volumes to be moved, thereby increasing privacy, reducing latency, and improving quality of service.
No longer relying on a central core also means the removal of a major bottleneck and potential single point of failure. Edge processing is based on distributed resources that may not be continuously connected to a network in such applications as autonomous vehicles, implanted medical devices, fields of highly distributed sensors, and a variety of mobile devices. To make use of AI in this challenging environment, an agile application that can retain learning and apply it quickly to new data is necessary. This capability is called inference: taking smaller chunks of real-world data and processing it according to training the program has done.
For inference to work in edge environments, processing architecture and hardware are required that are optimized and come with certain requirements on processing capacity, energy efficiency, security, and connectivity. NXP has established leadership in machine learning at the edge – particularly for the inferencing tasks. The NXP portfolio covers almost the entire MCU & application processors portfolio that is used in modern AI applications: i.MX 6, 7 & 8 product families, Kinetis MCUs & Low Power Cortex, QorIQ® communications processor portfolio, and S32 MCUs and Microprocessor Units. In fact, we have been ranked as one of the world’s top three artificial intelligence chipset companies.
To build innovative AI applications with cutting-edge capabilities, developers depend on a machine learning software environment that enables easy integration of dedicated functionalities into consumer electronics, industrial environments, vehicles, and other embedded applications, in general. But to roll out AI-based business models at a much broader scale and to make AI applications available to billions of end users across all verticals, the industry must first overcome past limitations.
This is why NXP has developed machine learning hardware and software, which enables inference algorithms to run within the existing architecture. The NXP ML environment enables fast-growing machine learning use-cases in vision, voice, path-planning and anomaly detections, and the integration of platforms and tools for deploying machine learning models, including neural networks and classical machine learning algorithms, on those engines.
And this is only the first step, as NXP is already working to integrate scalable artificial intelligence accelerators in its devices that will boost machine learning performance by at least an order of magnitude.
By designing things with smart properties and connecting them into the Internet of Things, we have created a global web of assets that have enhanced our lives and made them easier and more secure. The IoT gave us eyes and ears, and even hands, to reach out from the edge of the network into the physical reality where we gather raw information, which we stream to the cloud, where it’s processed into something of superior value: applicable knowledge.
By adding high-performance processing, we’ve started to process and analyze information less often in data centers and the cloud, and more now at the edge, where we see the magic occur. We witness that magic in smart traffic infrastructure, in smart supply chain factories, on mobile devices, in front-end stores, and in real time, where all the action takes place that makes our lives colorful.
The IoT in its present shape has equipped us with unprecedented opportunities to enrich our lives. Yet, it is only a stopover on the way to something even bigger and more impactful. We are talking about the artificial intelligence of things.
Today’s smart objects, even though they stream data, learn our preferences and can be controlled via apps, they are not AI devices. They ‘talk’ to each other, yet they don’t play together. A smart container that monitors the cold chain of a supply of vaccines is not an AI system unless it does ‘something’, such as making a prediction about the temperature development in the container and automatically adjusts the cooling.
An autonomous car or a search-and-rescue drone that autonomously navigates off-shore is in fact an AI system. If it drives or flies on behalf of you, you can trust that some serious AI capabilities are involved. Reading, speaking or translating language, predicting the mass and speed of an object, buying stock on your behalf, recognizing faces or diagnosing breast cancer, are all artificially intelligent characteristics when done by an algorithm.
Now, imagine a world in which the entirety of AI things was connected. Expanding the edge of the IoT with cognitive functions such as learning, problem-solving and decision-making would turn today’s smart things from mere practical tools into true extensions of ourselves, multiplying our possibilities to interact with the physical world.
As an integral part of the IoT, artificial intelligence is the foundation for entirely new use cases and services. Siemens®, for example, is using AI to improve the operation of gas turbines. By learning from operating data, the system can significantly reduce the emission of toxic nitrogen oxides while increasing the performance and service life of the turbine. Siemens is also using AI systems to autonomously adjust the blade angle of downstream wind turbines to increase the plant’s yield. GE’s drone and robot-based industrial inspection service, Rolls-Royce® IoT-enabled airplane engine maintenance service and Duplex’s AI voice are other examples to prove the move towards the AI of things.
The truth is, even with a broad range of nascent AIoT applications emerging, we can’t even fathom what else is coming. One thing is for sure, though – today’s digital age society is undergoing a fundamental change. The paradigm shift that comes with the convergence of AI and the IoT, will be even greater than the one we have witnessed with the introduction of the personal computer or the mobile phone. NXP is driving this transformation with secure, connected processing solutions at the edge, enabling a boundless multitude of applications in the future AIoT.
Continue reading part II of this blog series under this link to learn about AI implications for IoT Security.