The pace of edge computing is advancing so fast that today’s “breakthrough” is eclipsed in a short time by another, and then another. In this environment, it’s not easy to extract one or two things that will drive this technology, but I want to highlight two important factors: machine learning at the edge and the need for a common code of AI ethics to provide trust as AI becomes a ubiquitous part of edge computing.
Only recently, data from edge nodes was aggregated in a gateway and sent to a cloud data centre where massive computational resources could tackle ML training and inferencing. With the expansion of the smart edge in thousands of Industrial and IoT applications, it became clear that the latency, energy, and always-on connectivity required to send data to the cloud for processing and get a return response was untenable for most edge applications.
A better approach was to split the processing and analysis functions between the cloud and the edge itself. Thus, edge computing arrived.
However, with the growing complexity of the applications, the smart edge had to evolve – from smart to intelligent.
The intelligent edge is not just about fast response time but is also about training the edge to take decisions locally based on a set of real-time data, without the aid of the cloud or the gateway. Of course, systems could be built such that the edge reaches out to a local gateway or the cloud in case it is presented with a scenario it is not trained for, but such situations by design are expected to be very few in an intelligent edge network.
In a constantly evolving world, a true intelligent edge should have the means to evolve as well, so as not limit the lifetime of the product in the field. It is imperative that product developers think about the future possibilities and build in a balanced set of attributes to enable in-field enhancements. That could come in couple of different ways: having sufficient compute and memory capability and a way to receive the updates. This benefits both consumers and the manufacturers - consumers can receive the latest features delivered on their device and manufacturers can extend the lifetime of the device as well as build brand loyalty and revenue streams through additional services and support.
Many of these intelligent edge devices are in resource-constrained environments, where they must compute and communicate using little energy.
Additionally, most edge application solutions must be inexpensive, frugal with power, typically operating from tiny rechargeable batteries, possibly using energy harvesting. Adding ML functionality would make them power-hungry and too expensive to be practical. There were doubts if the required intelligent performance could ever be achieved with a small and inexpensive device.
Today, microcontrollers and crossover applications processors have blossomed. In our view, this represents what will ultimately become a new domain, where machine learning is performed right at the edge, consuming milliwatts of energy while delivering truly remarkable compute performance.
This is not a pie-in-the-sky prediction; but such highly capable MCUs and crossover applications processors are available now and are making their way into thousands of applications. The next generation of these edge processors will have dedicated neural network processing engines to provide astounding ML inferencing performance while also improving energy efficiency.
An MCU-based edge solution can now perform tasks such as secure face recognition using tiny, inexpensive IR sensors that are difficult to ‘spoof’. This is achieved through the use of ‘liveness’ detection in which a ‘warm body’ with precise characteristics is sensed, differentiating a human face from even a high-resolution photo or mask, all within the span of hundreds of milliseconds with a false-positive rejection accuracy approaching 99%. A high false-positive rejection accuracy is required in cases where security is paramount, be it the front door of a smart home or access to heavy machinery in an industry, where granting access to the wrong person because of incorrect detection (false-positive) is more dangerous than rejecting a genuine user (false-negative).
This brings us to the next issue we would like to see gain more attention in 2021: ethical AI.
As processors bring AI and ML to more applications, the issue of endowing the technology with a moral compass becomes not just a theoretical conjecture but a here and now reality. The problem, fundamentally, is that ethics is subjective with widely varied views, driven by geography, culture, values, and governments, thereby making it challenging to “define” a universal set of AI ethical principles. This becomes even more challenging on a global scale, as cultures may have different views on what is acceptable and what is not.
Fortunately, this issue has not been lost on the tech industry or by several governing bodies around the world, which are taking steps to develop policies and laws focusing on a human-centric approach to AI ethics. These initiatives envision ethical codes of AI conduct based on the principles of transparency, fairness, safety, and privacy. If these initiatives can be harmonised into a universal ethical code and adhered to, then
people can be assured that the AI/ML will act to enrich lives - by making bias-free decisions, by protecting personal privacy, and by ensuring that lives will not be invaded by companies, industries, and governments without consent. This is not a trivial task, but is required if AI/ML is part of our future. Without these guiding principles, there may come a time when the public at large will no longer accept AI. Therefore, we believe attention to ethical AI will be one of the most important developments to watch in 2021.
At the hardware and software level, ethical AI/ML cannot be achieved as a “bolt-on” extension, but rather must be thought through and implemented from the inception of a system. That includes high-level, broad-based security that is grounded in the silicon, software layers that add additional levels of security and AI/ML applications that conform to the ethical principle of bias-free training and essentially obey the Asimov’s first law (paraphrased for ML): ML may not injure a human being or, through inaction, allow a human being to come to harm. This means controlling action a device can do to prevent commitment of a moral crime; autonomous judgement to not cause injury – e.g. knowing when to stop a task if injury to humans is imminent; and including privacy indicators for transparency of what function a device is performing – e.g. listening-only mode or recording mode, etc. that are made clear to humans.
In summary, for industrial and similar applications, moving machine learning into the edge means better performance, while for people it brings AI closer to their personal lives. So, practically speaking, intelligent edge compute and ethical codes for AI are not as disparate as they seem but are joined at the hip, or rather at the edge, where the data is generated.
Author details: Ron Martino, senior vice president and general manager, Edge Processing business at NXP Semiconductors