Accessible on-device ML with Google Coral

The goal behind the platform is to solve practical problems with easy-to-use tools.

The idea behind Coral

Coral was created to address the growing need for machine learning on Edge devices. With more smart devices and IoT systems in use, the demand for solutions that process data locally, rather than relying on the Cloud, surged. Coral’s platform was created to enable developers to move from initial concepts to scalable deployments with minimal friction, and its ability to process data on the device means that it performs faster, has better privacy, and it can operate in environments where internet connectivity is limited or unavailable.

Supportive hardware

Coral offers a range of hardware options for developers such as the Coral Dev Board, which is a single-board computer that includes machine learning capabilities and supports various sensors for prototyping. Then there is the USB Accelerator, which is a compact device that plugs into any Linux-based system, including Raspberry Pi, to add ML processing power.

At the core of Coral’s hardware is the Edge TPU, a custom chip designed for efficient ML processing. It supports TensorFlow Lite models and delivers high performance while using very little power. In addition to this, tools like the PCIe Accelerator and Coral Accelerator Module make it easier to integrate Coral into custom designs, while the smaller, energy-efficient Dev Board Mini is ideal for more compact projects.

Easy-to-use software

Coral’s software is designed to work in tandem with its hardware. Mendel Linux ensures smooth integration, while APIs in C++ and Python give developers straightforward tools to access hardware features and deploy machine learning models. The platform supports TensorFlow Lite and offers pre-trained models for tasks like object detection and classification, and developers can also use transfer learning to customise models for specific applications.

The platform has also updated its offerings to improve the development experience, including model pipelining for distributing tasks across multiple TPUs and a model partitioner for better performance. The library of pre-trained models has also expanded to include MobileDet, a lightweight object detection model for mobile and Edge devices.

Real-world applications

Coral can be used in a variety of industries such as in agriculture, healthcare, smart cities, and environmental monitoring. Companies are also using Coral to prevent water loss with smart meters, while Google’s Series One meeting kits use Coral for advanced noise cancellation in video conferencing.

Making AI accessible

The goal for Coral is to build a community by making machine learning tools easy to access and use. Partnerships with organisations like PerceptiLabs and SigFox are helping to grow a community of developers who can build and share new applications. Free resources, including PCB designs and pre-trained models, make it easier for anyone to get started with Edge AI.