The history of computing and computers goes ages back. Which led to the development of electrical computers as we know it in the 20th century. Our first home computer was in the attic and mainly used for playing Pacman. That the personal computer would change our lives and the way we do things was hard to imagine. The same goes for my first mobile phone. It was not the first mobile phone ever, but still one of the first. I mainly used it for the game snake. Now here we are, many years later, the mobile phone became smart, and is indispensable in our daily lives. Who could have ever imagined the impact of home computers and smartphones? It doesn’t end there. Many devices around us are smart, collect data, and help us with useful actions and insights. Devices such as Google Assistant, Alexa, and Fitbit are already omnipresent. How does the playing of games lead to the fast rise of smart self-learning devices? In this blog, I will explore the major shifts and the rise of Machine Learning, and Tiny Machine Learning. Will this also start as a fun game and end up changing our lives?
Shifts in capabilities and capacities
So, our home computer started in the attic. Rapidly, after major updates, improvements in memory, and applications, it won itself a prominent place in the living room. Text processing programs, spreadsheets, and later also home baking made the computer very useful. With the start and development of the ever-expanding internet, possibilities became endless. Sensors, like a camera and movement sensors, GPS, for example, introduced its way into our lives. Currently, only 1% of the generated data on the edge is being analyzed. Can you imagine how much insight and intelligence we would gain if computational power would be in more devices, and things around us? Not only in our homes but also our offices, factories, healthcare, and agriculture, this impact will be enormous.
From the cloud to the edge
Nowadays most computational work is done in the cloud. Which means all devices send data such as photo’s, video’s, and texts to the large data centres of companies such as Google. This process takes a lot of time, energy (battery), bandwidth, and money. There are severe privacy risk and safety risk in this process. To overcome these problems, there is an increasing number of possibilities. Techniques to compute on small devices have been available for a few decades. However, techniques, battery life, to make such small devices which can run and store enough to compute weren’t there yet. Fortunately, now they are. Can you imagine the smallest Microcontroller is smaller than a golf ball dimple? This means an opening for a whole new world of possibilities. Is the future tiny, bright, and tiny ML?
The future is Tiny ML
So, let’s assume that a large part of the cloud shifts to the edge. Which I find plausible, since the technical possibilities, and the number of problems which could be solved using Tiny ML, programming alone could not solve which. There are privacy, speed, and cost aspects. If small devices, and things, could run their algorithms and learn from their surroundings. This would mean low power and low-cost intelligence on the edge. This will enable machine learning right at the boundary of the physical and digital world. Then we would have even more power, data input to make algorithms that will learn to solve problems by analyzing data for patterns.
Tensor input for Tiny ML
The data input of Tiny ML is like that of other forms of Artificial Intelligence. Tiny ML is a form of deep learning, which is Machine learning, which all falls under Artificial Intelligence. The data which is generated comes from sensors. Such as cameras, voice, temperature, light, smoke, gas, alcohol, pressure, accelerometer, heart rate sensors, and so forth. In this way, the device can sense its environment. With Machine learning, it can analyze what is going on, and the device can act upon it, all in one device.
The philosophical transition
Every cardinal technological innovation, which changed the way we do things, brings a significant change with it. Not all technological inventions change our lives drastically or force a paradigm change. A paradigm is a set of conventions and assumptions within science, which dictates how things are done. The discovery of fire, and the invention of the wheel, for example, are game-changing discoveries and inventions. The invention of the wheel was the prerequisite for everything which moves on wheels. Before the invention of the wheel, there was simply no world thinkable in where you would drive the car. The same goes for the possibilities of Tiny ML, the ability to deploy Machine Learning on small devices opens a whole new world with opportunities, possibilities, and threats. The threats I will discuss in the next section.
Ethics: responsible AI
According to Vijay Janapa Reddi, Associate Professor in the John A. Paulson School of Engineering and Applied Sciences (SEAS) at Harvard University, Tiny ML is disruptive. There are a lot of benefits, but also serious risks, and there are going to be some serious ethical challenges we need to be concerned about. There is always that people will use the technique for bad purposes. This is called multi-stability. This is the reason Google holds off its facial recognition technology. We should also consider conflicting values by different stakeholders. According to Susan Kennedy, she has a PhD in philosophy; it is important to use a human-centred design framework. It is important to keep the stakeholders and potential uses technology in mind at the earliest design stage. And so it does at the stage of development and deployment. This is extra important since once you deploy, it is hard to go back. The debug function is often absent.
Benefits of Tiny ML
Many processes will become cheaper, faster and closer to the place where data is generated. Humans will get more streamlined product design and processes. We will get smart offices, smart factories, smart homes, smart cities, smart public space, and smart healthcare. There will also become a better connection between the analogue and digital world. Where we will use the best of both worlds. The benefits of an analogue neural network in lower power. That will mean that more intelligent mobile devices that run longer on a single charge. The possibilities are endless, and they estimated it that by 2025, 75% of enterprise data will be processed at the edge. Compared to only 10% today.