They say good things come in small packages, but what comes in tiny packages? According to Harvard associate professor Vijay Janapa Reddi, who also helped found the nonprofit machine learning (ML) organization MLCommons, tiny machine learning, or TinyML, has the potential to be really big, really soon.
“The end user value with TinyML is pretty remarkable,” said Janapa Reddi, who teaches and helped shape the HarvardX Professional Certificate program on the subject. “Because you can get this technology into the hands of individuals so cheaply with all kinds of open source software out there, and then put open source machine learning methods on top of it, you can enable a lot of intelligence on these devices. That's really why there's so much excitement about TinyML.”
Wondering if (and how) TinyML could give your career a big boost? Let’s begin by understanding what it is, why the time is ripe for it, and where learners can go next with a certificate in TinyML.
OK, Google: What Is TinyML?
An easy rule of thumb on the consumer side of things: If it has a wake word (i.e., it responds to “Alexa,” “Hey Siri,” “OK Google,” or similar), it’s using TinyML.
Whether it’s a consumer product like Google Home or an industrial tool intelligently monitoring systems in real time, anything that’s “smart” has to get its smarts from somewhere. Machine learning is how they do it. What makes TinyML so relevant is that it focuses on small, low-powered microcontroller devices that have been proliferating in everyday life, as well as in the industrial and logistics realm.
Microcontrollers are small, specialized computers within a larger device, tasked with a single program or mission such as changing the TV channel based on input from the remote. Printers, cars, medical devices—almost every modern electronic device is housing a microcontroller.
TinyML combines machine learning with embedded systems to enable these tiny devices to operate independently at the point of data collection, rather than running back to the mothership with each and every request—so when you ask the device a question, it can answer right away using local functions and very little power. This on-device machine learning is fundamentally different from running machine learning in big data centers or on the cloud.
Consumer and Industrial Examples of TinyML
On the consumer side, the power of TinyML has largely been restricted to the personal assistants world, where we’re familiar with using wake words today, but more microcontrollers means more possibilities. Smart appliances, for example, are a clear opportunity, enabling intelligence in devices such as ovens, where a “bake at 350” command could automatically turn the machine on and set the temperature.
An example on the industrial side is predictive maintenance; the ability to know if a machine will fail ahead of time.
“For that, you have to collect data in real-time, right at that sensor point, and the aggregate sensor bandwidth is huge because you're going to have so many of these sensors, each one able to generate a lot of data. There is no way you can pump all this data back up into the cloud,” Janapa Reddi said. “With these small, smart electronic sensors, if we can just embed them everywhere, then they would be able to raise alerts proactively, which means less downtime and much more efficient manufacturing pipelines."
In short, TinyML technology is not reserved for just one purpose or type of device. The possibilities are endless, and that means the opportunities are, too.
Why TinyML is About to Get Big
These are the critical years for TinyML, as the cost and accessibility of the hardware and software ecosystem has matured to the point of democratization that will take it to the next level.
In the evolution of smartphones, this was the stage where we began to see innovation show up on an individual level, with everyday people leveraging the app store to bring their own ideas and tools to fruition. Janapa Reddi predicts the same will happen with TinyML very soon. Those who get into this space now will have the unique opportunity to help shape this industry in real-time, as it is being formed.
With rapid growth propelled by the burgeoning Internet of Things, there are already 250 billion microcontrollers in use today. That’s like every human on earth owning 32 Apple Watches. In addition to being a funny visual, it’s truly a staggering number of devices—and the growth rate is only increasing. Janapa Reddi projects that the world will reach 1 trillion microcontrollers (or 130 Apple Watches per person) a lot sooner than we realize.
Janapa Reddi says that’s because the versatility of TinyML gives it some immunity to the “skepticism” phase of the hype cycle. His prediction: “I don't expect that the dip is going to be as large as you typically see with extreme cutting-edge technologies like autonomous driving or AR and VR, which are significantly harder to implement on a mass scale,” he said. “TinyML is not going to go through that sort of dip.”
What Can You Do With a TinyML Certification?
According to Janapa Reddi, industry professionals want new hires who specialize in one area but understand the big picture. Studying TinyML is a great way to gain that perspective. The HarvardX Professional Certificate program Janapa Reddi helped create is designed to help anyone build these skills, from learning the fundamentals of machine learning, deep learning, and embedded devices to gaining experience using production scale software frameworks.
Combining computer science with engineering, the first-of-its-kind program—a collaboration between expert faculty at Harvard’s John A. Paulson School of Engineering and Applied Sciences (SEAS) and innovative members of Google’s TensorFlow team—prepares learners to tackle real-world challenges around TinyML deployments.
For those who are already working in the field, TinyML can broaden their professional capabilities and give them greater clout to stand out in the workforce, or even prepare them for a career shift that may not be possible with a more rigid skill set.
Whether you’re just getting started or trying to up your professional game, Janapa Reddi says it’s a mistake to simply study machine learning in terms of the job market, as this can lead to a siloed career and make it difficult to grow beyond a specific role. Studying TinyML is a more holistic end-to-end approach and can help provide an edge in the workplace, where those who understand TinyML are able to deliver greater value through versatility and innovation.
Why Anyone Can (And Should) Learn About TinyML
The TinyML Professional Certificate program on edX starts from square one, so those without a background in machine learning still have an onramp to learn and succeed.
The goal, said Janapa Reddi, is to provide the right building blocks to create an end-to-end application with real-world use cases while leaving the door wide open for each student to bring their own ideas and creativity to the table. Therefore the courses are not exclusive to current engineers or data scientists, but rather encourage a variety of students to enter the space with unique perspectives and innovations.
Equally important, he says, is for those outside of the discipline to grasp the technologies of TinyML, even on a basic level—just like everyone has to learn math. TinyML is “going to be so deeply ingrained into every single part of our lives that it should not be restricted,” says Janapa Reddi. “AI is one of those technologies that you do not want stratification. You simply did not want those technologies to be reserved for some elite or some subset of people.”
Learn more about HarvardX's TinyML program or explore other data science courses and programs on edX.