A practical approach for engineers.
By Ehsan Kourkchi, Data Scientist at SensiML.
Generative artificial intelligence (GenAI) has evolved from content creation to enhancing machine level (ML) model development for IoT devices. SensiML’s tools leverage this technology to streamline the process and build more efficient edge AI solutions.
When GenAI was first introduced, it was primarily used for content creation. However, it is now being adopted much more broadly, including being used to enhance the design and deployment of ML models for sensor-equipped IoT devices. SensiML’s Piccolo AI, an open-source AutoML solution, when combined with SensiML Data Studio, an ML DataOps application for managing ML train/test datasets, integrates GenAI into the workflow, allowing engineers to build more accurate and efficient ML models for edge IoT embedded applications.
Building production-quality ML sensor models for use on embedded and IoT edge devices presents unique challenges. Fortunately, GenAI and AutoML—as found in innovative software tools for the IoT edge like those offered by SensiML—can help to overcome them. Below, we touch on some of these and highlight one such feature available today in SensiML Data Studio.
Edge IoT AI development challenges
Developing ML models for edge IoT devices presents several key challenges, including the following:
• Collecting enough high- quality data for model training: A time-consuming and expensive task, especially for rare events such as system failures.
• Overcoming hardware limitations at the IoT edge: Edge devices have limited computing power, memory, and battery life, making it difficult to run complex ML models efficiently.
• Acquiring data science expertise: Choosing algorithms, tuning hyperparameters, and managing data pipelines requires specialized knowledge that engineers working on edge IoT applications often lack.
Edge IoT AI development solutions
GenAI and AutoML technologies can help address the aforementioned issues as follows:
• Techniques like quantization and pruning can reduce model size without sacrificing accuracy, ensuring they run efficiently on resource- constrained devices.
• Generative models can create realistic synthetic data to fill gaps, enabling training on a wider variety of scenarios, such as simulating rare failure events for industrial IoT.
• AutoML tools like SensiML’s Piccolo AI utilize machine learning methods in the process of building ML models themselves. In this way, they automate tasks such as feature selection and model tuning and optimization that would otherwise require human expertise in data science. This streamlines development and reduces the need for data science expertise.
Application example
Voice-enabled IoT applications readily lend themselves to the application of today’s GenAI technology. SensiML’s Data Studio now incorporates GenAI to provide an automated text-to-speech feature for rapid synthetic voice data development for training a voice recognition model. This approach provides a diverse set of voice samples, crucial for applications like smart home devices or wearables that rely on accurate voice control.
Using GenAI, engineers can easily create training data to account for different voices, accents, and speech patterns, improving model accuracy for wakeword detection, voice command and control, and voice authentication applications. With SensiML’s Piccolo AI, this process becomes more streamlined. Automating the data augmentation allows engineers to develop edge AI systems capable of accurate voice recognition without needing to collect an exhaustive set of real- world voice samples.
Summary
GenAI and AutoML are making edge AI development for the IoT more accessible. They address challenges like limited data and hardware constraints, allowing engineers to create efficient ML applications. Tools like SensiML’s Piccolo AI simplify the process, and GenAI makes IoT development more scalable, meeting
the demands of modern sensor-equipped designs.