SensiML provides end-to-end AI toolkit for SensorTile.box

This article is compiled from ST official blog

Is there a way to simplify the creation and validation of machine learning algorithms at the edge? This was a popular talk from SensiML during the 2020 ST Virtual Developer Conference. AI had a major presence at the show, and during the discussion, panelists explained how customers are starting to ask for products with machine learning capabilities. So we thought it would be worth sitting down with Chris Roger, founder and CEO of SensiML, a member of the ST Partner Program, to learn more about the tools the team presented at the ST Developer Conference and their importance.

Challenge: Front-end data collection

Although experts have been discussing it for years, machine learning at the edge is still a young technology. As a result, engineers often have to rely on many different tools and complex workflows. As a result, they may face greater challenges when teams move from one step to the next. SensiML solves development challenges with an end-to-end toolkit that includes data capture, algorithm generation, and validation. In addition, the SensiML solution is also transparent and can be extended through a GUI and Python IDE. Therefore, even if the team has mastered machine learning, they can still take advantage of the toolkit. In fact, the company provides many tutorials, including video tutorials. Therefore, we thought it was necessary to push the discussion further and decided to talk to Chris about some of the main shortcomings of TinyML applications and what his team is doing to further optimize this workflow.

data collection

Warehouse and sensor data SensorTile.box support

Data capture itself is a difficult problem to solve because those who don’t have data have to make a significant investment to get it, and those who have data will certainly keep it to themselves and are reluctant to share it. Therefore, SensiML is a unique partner because the company solves both problems. First, it provides a data warehouse that serves as a repository for datasets. Some examples even come with tutorials to help teams that are new to machine learning applications. Second, the SensiML toolbox connects to the SensorTile.box and utilizes ST’s sensor platform to collect information. With just a few clicks of the mouse, the SensiML software can transmit the data collected by the sensor through the serial port.

From BLE to MQTT

As Chris explained in the talk, “The ability to quickly connect the SensorTile.box and immediately capture data is the result of a collaboration between SensiML and ST. SensiML started developing the first version of the SensorTile and received feedback from the ST team. Our engineers then worked on the SensorTile.box and the STWIN sensor kit. We will also support the LSM6DSOX and its machine learning core by sharing simple decision tree models that show different ways to wake up the MCU.”

Chris also described how engineers used SensorTile to evaluate the limitations of BLE. The low data rate of the protocol forced them to rely on external memory cards because the system could not transmit all the data at once. As a result, the SensiML team realized that they needed to adopt more powerful transmission technology, especially for industrial applications. Chris explained that his team worked on using the MQTT protocol for transmission with ST sensors, which took time to debug, and they ultimately used MQTT over serial to reduce overhead while benefiting from the higher data rate. Today, SensiML users can take advantage of all the features with just one option.

Optimize workflow

Automatic Tagging

Another reason the SensiML presentation was so compelling is that it addresses a challenge that’s often overlooked by designers: labeling. Teams with one or more AI projects know how difficult data collection can be. We recently highlighted this point when we shared how ST engineers developed a crying baby detector app. Data acquisition is challenging. However, once engineers have a dataset, they still need to label it, which is another complex process with a huge impact. Most competing frameworks start with huge CSV files that are unwieldy, so the labeling process becomes cumbersome.

SensiML solves this problem with tools that catch signals and label them. For example, its Data Insight technology automates labeling operations. Users first describe dozens of examples, and then the toolbox infers what it thinks are the same samples. As a result, engineers only need to verify the tool's automatic labels or change some false positives. Ultimately, the process becomes simple, fast, and enjoyable, and SensiML also provides version control capabilities. If a team works on multiple datasets and wants to go back to previous annotations because of a problem, they can use the rollback function to go back.

Using STM32CubeMX and LSM6DSOX

In addition, Chris also told us that the next step in workflow optimization is to better interact with STM32CubeMX.

He said: “SensiML is working on an interface with ST configuration tools. Our software works well with SensorTile.box, but STM32CubeMX provides us with a variety of customizations and development boards. Our goal is to make the program part of our workflow so that users spend less time in it. We have a series of files for sensor and MCU configuration in SensiML. Our team aims to no longer have users generate them manually, but to do part of the work for them to speed up the customer's prototyping phase.”