Making technology compact is one of the focal areas of modern science and technology. Miniaturization played a crucial part in making technology become more easily accessible to its users, significantly in the past couple of decades. The computers that once were the size of a small bedroom are now thousand times powerful and smaller. The trend of miniaturization bolstered growth of many latest technological advancements. One such technology is TinyML (Tiny Machine Learning), the fastest growing field of machine learning technology. It is no news that we are gradually ushering into a world where AI accounts for most technological advancements and has paved its way into our daily lives. Some easy examples are: voice recognition on our smart devices, real-time commute predictions on Google maps, facial recognition systems and much more.  

Ever wondered what happens or is the behind the scenes when you say Ok Google, Hey Siri or Hey Alexa? Well, that’s precisely TinyML in action. This blog discusses about how TinyML makes edge AI a reality and ubiquitous. Also, you’ll learn what is TinyML and the tiny revolution it is about to make in the realm of artificial intelligence.  

What is TinyML? 

TinyML or Tiny Machine Learning is a subset of machine learning that shrinks deep structured networks to fit on tiny devices, primarily optimizing machine learning models to run on microcontrollers, digital signal processors or devices that are cheap and power-constrained. Research indicates that there would be around 2.5 billion devices that will be equipped with TinyML technology in the market by 2030. Simply put, TinyML is machine learning running on low-power devices without connecting them to the internet.  

TinyML in action 

Tiny Machine Learning. Where did this idea originate, and who did it first? Well, it was Pete Warden, widely considered the father of TinyML, who came up with the idea of TinyML when he learned that the Google OK team was able to run voice interfaces and wake words with only 13 kilobytes. Later, he developed this idea to enhance the way machine learning can be utilized on low-powered devices. 

“TinyML will be pervasive in many industries. It will impact almost every single industry: retail, healthcare, transportation, wellness, agriculture, fitness, and manufacturing” says Pete Warden, the technical lead of the TensorFlow Micro team at Google. 

It is evident that voice/virtual assistants like Siri, Ok Google, Alexa have become a part of our daily lives, from helping us look up on the things we are curious about to guiding us to a location on Google Maps to facilitate hands-free entertainment while watching TV or listening to something on Spotify. However, the process is kind of complicated, as these voice assistants rely on certain wake words to seamlessly deliver the most appropriate response to users. It stands to reason that virtual assistants need to be constantly listening for their wake words. We all know what happens next. Yes, too much power consumption which results in poor battery performance. 

Thus, to make virtual assistants consume less power, Google decided to add an additional component, a tiny microcontroller, to its smart phones. As this microcontroller would only listen for the wake words, the voice assistants wake up automatically to the wake words and process the user’s voice command without relying on the processing power of the phone. Surprisingly, it only takes a 14 kB tinyML model that runs on a Digital Signal Processors. 

Advantages of TinyML  

As machine learning and AI need a massive amount of potential, structured data, which, in turn, would make the devices consume a lot of processing power, it is not advantageous for machine learning models to run on power-restrained devices. This is where the TinyML comes into play. Look at some of the many advantages of TinyML below: 

  • Low power consumption – One of the substantiate advantages of TinyML. As micro-controllers consume very little power, devices can be run for an extended period without needing to be worried about power consumption.  
  • Low latency and internet bandwidth – Since TinyML run on edge devices, there is no need for the cloud where the data is usually sent to run inference. No data sent to a server to run inference means faster inference with low latency. 
  • Privacy – Since TinyML models do not send data to servers or data centers, the risk of crucial, sensitive data being compromised is zilch. 

TinyML: the next BIG thing in AI  

According to the Emerging Spaces review of Pitchbook, ever since 2020, $26 Mn have been invested in TinyML. The rise of TinyML is about to create a paradigm shift in ways how the intelligent industry can harness AI and machine learning to the hilt, as TinyML runs locally on power-restrained devices. Industry 4.0 is all about digital transformation, and AI is at the heart of digital disruption. TinyML offers many unique solutions by summarizing and analyzing data at the edge on low power devices.  

Wrapping it up 

With TinyML  and its ability to run machine learning models on ultra-low powered for weeks, months and even years without the help of the cloud or internet, it is highly likely that AI becomes more ubiquitous and cheaper for organizations and their users. As we speak today, most AI applications and models across the world are predominantly relied on the cloud. With the emergence of embedded-AI and tinyML, there would be more privacy, as edge-AI systems are less vulnerable to cyberattacks and data breach.

Latest posts by Gokuladhasan Ramani (see all)