Felix
Google launched the Teachable Machine (TM), an AI software tool in a web version in 2017, followed by Teachable Machine V2 coming out in 2019.
How to get started?
Visit the TM website and click “Get Started”, and you will see the new project selection page, where users can choose to start three different types of projects- Image, Audio, and Pose. It is recommended that beginners start with Image Projects- the most basic ones.
In our cases, we use about 300 photos taken for each class decided. After collecting the image samples, the training started.
Parameters in the deep learning model
Under the “Training” icon, there is a button “Advanced” for a dropdown menu to show additional hyperparameters for different settings, which are commonly used for tuning in training neural networks.
In addition to the adjustment of training parameters, there are also data regarding the training process for users’ reference.
The training model splits all types of data into 85% for training and 15% for testing. The 15% will not be fed into the neural network, but will only be used to test the training results after the training is completed.
The two graphs below show the changes in accuracy and loss after each training. In theory, it is better to have values with higher accuracy and lower loss. These two graphs will come in handy for a complete deep learning framework, so as to tell whether the model can effectively converge.
Project export and saving
Makers all look forward to using the trained neural network model in different fields! After the training is completed, click “Export Model” and users will see a pop-up screen on the right with three tabs at the top for three different formats- Tensorflow.js, Tensorflow, and Tensorflow Lite. In the Tensorflow.js tab, users can choose to upload the completed model to Google Cloud or download it to a personal computer. At the bottom of the page are also code snippets for developers to apply their models directly to their web pages.
Shared projects
As TM is easy to use, there are several shared projects using it TM on the Internet. For instance, Tiny Sorter, integrated with Arduino functions, can tell cereals from marshmallows through simply a laptop webcam; the two different items can be put into two bowls with the use of servo motors controlled with Arduino systems. In this way, TM has evolved from an analyzing tool for software to one that can sort real-life objects!
Conclusion
Even though TM features basic functions and train classification models, there have been practical applications in education both at home and abroad. Quite often, what people need may not necessarily be a perfect and powerful Swiss knife, but simply a small tool with just the suitable functions that can inspire ideas and even be extended for more usages! Food for thought- how many people are there around you that are now using TM for machine deep learning?
