When is TinyML too much? - Stacey on IoT | Internet of Things news and analysis

• Read original article here

When it comes to machine learning (ML), sometimes every problem looks like it needs a neural net, when in fact it just needs some statistics. This is especially true when it comes to running algorithms on microcontrollers and for industrial use cases such as predictive maintenance, according to Bernard Burg, director of AI and data science at Infineon.

I focus a lot on using machine learning at the edge — specifically the idea of running machine learning models on microcontrollers, known as TinyML — because there are clear benefits for the IoT. By analyzing incoming data where that data is created, engineers can reduce latency, lower bandwidth costs, and increase privacy while also saving on energy consumption. But one doesn’t always need TinyML. Sometimes using linear regression or anomaly detection will do.

Burg told me that so far he’s seen three broad use cases where local data analysis needs to take place, and not all of them require a neural net or true machine learning.

The first use case is handling time series data. Time series data is simply a time stamp and the measurement of something, be that temperature, humidity, light levels, or whatever. Sure, companies can build a neural net to analyze that data, but it’s not necessary and oftentimes isn’t accepted because it can be hard to explain how a neural net makes a decision.

“Those models are in competition with control systems, and people have been building those models for 50-60 years,” said Burg. “So when we come in as AI guys we have to be able to explain why they do what they do.” In that case, Burg argues that a linear regression is probably a better option because it’s both simple and explicable.

The second use case is focused on sound detection. As the incoming data usually looks like a wave pattern, the goal is to match that wave pattern to a wake word, or the noise of, say, glass breaking to the same sound in the environment.

“For those models, people go normally to the neural net because at the end of the day you may not be challenged by an expert wanting an explanation for why it works,” said Burg. He added that you could also do simple pattern matching, especially in the case of anomaly detection for noises or vibrations in an industrial setting.

The third use case is all about embracing neural nets and machine learning. That is computer vision, where a computer is evaluating data in multiple dimensions as it attempts to classify an object. While you can do edge-based object recognition if the sensor is only trying to alert for a limited set of two or three objects, most vision models are large and will run on more power-hungry devices.

So a sensor might be able to recognize a human, and then wake up a larger processor to figure out who that human is. According to Burg, neural networks are the best option for these use cases, but they aren’t models suited for microprocessors.

What I took away from my conversation with Burg is that in some cases data scientists are wielding machine learning like a hammer — and many problems look like nails. But shrinking models down to work at the edge is complicated and costly. And with use cases such as computer vision, it doesn’t actually work.

So Burg’s caution about electing to use machine learning when algebra and statistics will work, especially in conservative industries such as manufacturing, is worth remembering. Even for me, a veritable champion of TinyML.