Train ML Models Without Writing Any Code.
Google Creative Lab consists of a small group of ‘interdisciplinary thinkers and doers’. They come from a large variety of professions and collaborate on projects such as Teachable Machine which are usually the first of their kind.
Teachable Machine is an online tool that allows its users to simply train machine learning models without writing a single line of code. Users need to enter data samples for two or more things they want the model to be able to distinguish between. These can range from identifying objects in images, classifying different sounds or even the different poses someone is stood in. For example, in an image classification model that distinguishes between a pencil sharpener and an eraser, you would record a short video of each object and each frame in the video will be used as training data. Once trained, the website allows you to export the machine earning model you just effortlessly trained and use it in your own projects for free. That’s as simple as it gets.
A GIF showing the simplicity of the Teachble Machine website. Source:https://experiments.withgoogle.com/teachable-machine
Here’s how it actually works:
Google has recruited a technique called transfer learning to make this project easily accessible to everyone. Deep learning models require enormous resources and large datasets to train to an adequate level. Transfer learning is a technique where a base ML model is well-trained for general tasks and it is then re-purposed and used in more specific, related tasks. This transfers the knowledge gained in one setting to the new task which allows for rapid progress and great performance while requiring minimum training data and resources when modelling the second task!
The pre-trained base model being employed here is called MobileNet. MobileNets are ‘efficient convolutional neural networks for mobile vision applications’. The creatively named training data for this model is called ImageNet. MobileNets are a set of lightweight convolutional neural networks that were developed by Google to run efficiently on mobile browsers. The model has previously been exposed to thousands of different images during its training process (done by Google) so we can use the final trained model as a feature extractor for object detection (finding multiple objects in the same image) or image segmentation via the Teachable Machine website. Oh, and on a side note, the model is trained locally in your browser which is why the website alerts users not to change the tab it’s running on.
Try it yourself: https://teachablemachine.withgoogle.com/train
Sources:
Article thumbnail credit: becominghuman.ai
https://experiments.withgoogle.com/teachable-machine
https://github.com/googlecreativelab/teachablemachine-community/blob/master/README.md
https://machinethink.net/blog/mobilenet-v2/
https://arxiv.org/pdf/1704.04861.pdf
https://www.youtube.com/watch?v=kwcillcWOg0&list=PLRqwX-V7Uu6aJwX0rFP-7ccA6ivsPDsK5