At the conclusion of this tutorial, we will revisit the problem context - the envisioned uses, stakeholders, and values that guided our design and development.
Envisioned Use
Given the small dataset we used to train this model and its performance (often misclassifying still gestures) we recommend that users limit their use to fun, low-risk situations. Personal projects like games or home automation projects are ideal for this project.
We strongly caution against using this interface as a regular input for computers or machines. Since the model sometimes thinks a gesture is being performed when the user is still, overusing or using this project in a critical system could lead to misfires that cause frustration.
Stakeholders
We believe that people who can perform gestures in a standard manner have a lot to gain from using this project in their personal projects. However, anyone who cannot easily perform these gestures will not be able to make use of it. This is due to a failure on our part in not being able to collect diverse enough data. In future work, we will try and strengthen the dataset by including more people. We also encourage interested individuals to use Model Blocks to train a model on their own dataset.
Values
The primary values of this project are fair accuracy, working well for different people, and easy natural use. As long as this project is only used in low-stakes situations, we believe the model can be accurate enough. This model may not work well for everyone, we recommend that people test the model further with their own movements before using it. As we receive more information about how well this works for others, we will be able to improve the model greatly. Finally, we do believe that we chose a set of gestures that will be pretty easy and natural.
ShotSpotter is an example of a TinyML system that has been deployed in many large cities such as Chicago, San Diego, Cleveland, Atlanta, and Minneapolis since 2016. The technology detects the sounds of gunshots and then reports them automatically to police departments. When the system correctly detects gunshots, a quick response from police officers can help save the lives of gunshot victims and reduce violence in communities. However, false positives have also been a huge problem. Inaccurate reports result in police swarming to neighborhoods looking for an armed suspect, creating a tense situation for both the police and unsuspecting residents of the neighborhood. The true value of the ShotSpotter system lies in understanding the value of its benefits in the context of its potential harm.
This concludes this tutorial, but the development and improvement of your project is a continuous project. Complete an impact statement to document your project to share with others who will enjoy and may build upon what you have created.