Edge Impulse, a San Jose-based startup, aims to bring AI capabilities to microcontrollers that power consumer electronic devices. The company has built a cloud-based platform to enable developers to build AI models targeting microcontrollers.
Edge Impulse combines two popular techniques to bring AI to microcontrollers – AutoML and TinyML.
AutoML focuses on two critical aspects of machine learning – Data acquisition and prediction. The AutoML platform will abstract all the steps that take place in between these two phases. Essentially, developers bring their own dataset, identify the labels, and push a button to generate a thoroughly trained model that’s ready to predict.
TinyML is a technique that optimizes machine learning models for embedded devices that are resource-constrained. These embedded devices run on batteries with limited processing power and memory availability. Traditional ML models cannot be deployed on these devices. With TinyML, the models are converted and optimized to run on the smallest unit of an electronic device – the microcontroller.
MORE FOR YOU
Targeting microcontrollers for running AI is a complex and time-consuming process. Developers need to acquire data from various sensors attached to the microcontroller. Then, they need to pre-process and normalize the dataset before feeding that to a complex neural network. The artificial neural networks need massive computational power delivered by high-end CPUs and GPUs to train the models. Once the model is trained and evaluated for accuracy and precision, it is converted and compacted to run in microcontroller-based devices.
Edge Impulse managed the entire pipeline from data acquisition to model deployment. It has tools that can be run in Windows or macOS to collect data from sensors and ingest it into the cloud-based platform. Depending on the data format, such as time-series, audio, and video, developers can choose appropriate neural network architectures and optimization techniques to train the model. Finally, Edge Impulse converts the model, optimizes it and generates an artifact that can be directly deployed in the device. All of this is done without writing a single line of code.
Behind the scenes, Edge Impulse leverages the infrastructure, frameworks and tools typically used by expert data scientists and ML engineers. It takes advantage of parallelization for data processing and sampling the datasets. TensorFlow is used to build and train neural networks backed by powerful AI accelerators such as GPUs.
TensorFlow Lite, a flavor of TensorFlow optimized for devices, is used for converting and compacting the model. Edge Impulse can also generate a Web Assembly or a TensorRT model optimized for NVIDIA Jetson-based edge devices.
The key advantage of running AI in embedded systems is performing predictions without depending on the cloud or external services. The models become a part of the firmware flashed into the devices.
Edge Impulse has partnered with silicon companies and DIY electronic kit developers to accelerate building AI models for microcontrollers. It is gaining traction within the hobbyist developer community and IoT companies to bring AI into the embedded devices ecosystem.
Last month, Edge Impulse raised a $15 million Series A investment round led by Canaan Partners with Acrew Capital, Fika Ventures, Momenta Ventures, and Knollwood Investment Advisory. This funding will further the mission to democratize ML for millions of developers and engineers deploying to billions of edge devices.