This repository contains resources related to the machine learning (ML) sensor datasheet. The ML sensor is a new paradigm for edge devices which segregates the sensor input data and ML processing from the wider system at the hardware level while providing a thin interface that mimics traditional sensors in functionality. An in-depth description of ML sensors is provided here.
The ML sensor datasheet, as outlined in "Datasheets for Machine Learning Sensors," provides an overview of both the ML and hardware characteristics relevant to the device. This builds upon prior work focusing on specific aspects of the model pipeline (e.g., datasheets for datasets, data nutrition labels, model cards), and the existing industry-standard for providing a datasheet for hardware sensors. Related works are summarized in the table below.
The items listed within the proposed datasheet include but are not limited to: environmental impact, communication protocol, device diagrams, dataset attributes, model characteristics, end-to-end performance analysis, compliance and certifications, and hardware specifications. Including these details in an easy-to-digest datasheet provides many benefits as it provides transparency, explainability, and auditability. The ML sensor paradigm itself provides added privacy, security, and user-friendliness, which empowers end-users to more easily control their data, as well as for developers to more easily integrate ML capabilities into sensing devices without requiring ML expertise.
The proposed structure of the ML sensor datasheet is shown below.
A full example datasheet for a person detection ML sensor is provided here.