tinyML Talks on July 6, 2023 “Creating individualized solutions for industrial-grade and environmental problems with TinyML” by Kutluhan Aktar from Edge Impulse

We held our next tinyML Talks webcast. Kutluhan Aktar from Edge Impulse presented Creating individualized solutions for industrial-grade and environmental problems with TinyML on July 6, 2023.

In light of recent developments in artificial intelligence and machine learning, it has become more tempting to apply machine learning to solve industrial-grade, environmental, and health-related issues. Nevertheless, at least for now, devising a solution for a large-scale problem with ML can lead to interminable workloads and exorbitant costs depending on the nature of the targeted problem. For instance, building a ubiquitous pest detection system with object recognition requires an abundance of miscellaneous data sets due to differing subspecies, soil types, environmental factors, etc.

Instead of focusing on solving a problem in every possible scenario with ML, we can create individualized solutions for industrial-grade and environmental issues with considerably low budgets and workloads. Like individualized treatment plans, the accumulation of tailored and refined ML solutions for large-scale problems instigates a significant surge in revolutionizing our world. In this presentation, to fortify the concept of benefiting from TinyML and edge devices to create individualized solutions for large-scale problems, I will demonstrate some of my proof-of-concept AIoT projects.

I am a self-taught developer and maker who enjoys contemplating proof-of-concept AIoT projects in various fields. I was an aspiring physics major, but I decided to drop out of university in order to follow my vocation to be an independent researcher and build original projects from scratch. With the help of lots of innovative companies, I have been able to keep devising inspiring projects and realize my ideas in recent years as an occupation.


Watch on YouTube:
Kutluhan Aktar

Download presentation slides:
Kutluhan Aktar

Feel free to ask your questions on this thread and keep the conversation going!


Q: Can you programmatically upload data?
A: Since I utilized Edge Impulse to build and train my models, it is possible to upload new or existing data from any device via the official data forwarder:

Q: Is collecting/creating the word or data for analysis the largest/ most important part of the project?
A: Absolutely. Constructing a valid data set is the most crucial step for me while developing a project. I always create a unique data set for each project, even if I need to design a controlled environment from scratch. Algorithms and models are merely gratuitous tools without proper knowledge of the targeted problem (data).

Q: i want to learn tinyml. but i dont how to extract the right information . and i have some projects related iot
A: I highly recommend inspecting AIoT project tutorials on platforms like Hackster, Hackaday, and Instructables. Also, you can check the Edge Impulse GitHub repository for code examples:

Q: how start with ?
A: You can start with the Edge Impulse Expert projects, including some of my AIoT projects as well:

Q: How do you make the leap from the bench scale to commercial scale?
A: I build open-source proof-of-concept projects to fortify my approach to creating individualized solutions with TinyML. Therefore, I do not have much experience in commercialising my TinyML projects or their potential market success.

Q: exactly what I was looking for -practical ML approaches using off the shelf- technology
A: Thanks for your kind words and support :slight_smile:

Q: which sensor did you use for pipleline diagnostics?
A: I utilized the MR60BHA1 60GHz radar module provided by Seeed Studio:

Q: Great projects Kutluhan! What method do you use to validate the models to ensure they are accurate please?
A: Thanks :slight_smile: For most projects, I collect a separate data set and utilize Edge Impulse model testing to validate how well my model will perform on unseen data.

Q: very, good works. congratulations. i would like to know at which stage are you filtering sensor data noise, and which tools are you using for such purpose…
A: Since I usually utilize preconfigured sensors, I do not encounter faulty samples or data noise very often. Nonetheless, I always use Edge Impulse data explorer to visualize my data set and check if there is any problem with my uploaded samples.

Q: Thank you Kutluhan and Shiyun!
A: Thanks for your kind words and support :slight_smile: