I see MobileNet and ResNet embedded as application resources. Because of the speed of the training, I suspect they are doing feature transfer learning.
While the UI is quite nice (thanks to Mike Matas, I am sure), I don't see a strong advantage to using this on MacOs, when CreateML is available. CreateML doesn't have the simple interface of Lobe, but the UI is quite accessible and gives you access to additional classifies, like sound, text and tabular data. If you need ever more power, you can use TuriCreate if you want to stay in the Apple ecosystem.
The simplicity of the UI is a feature, but also a disadvantage when you start having more than a handful of labels and training images. I totally see how Lobe could be a nice intro into the world of labelling and classification.
The model training time using Lobe is comparable to similar transfer learning tasks using other machine learning frameworks. I did a comparison between Lobe and Turi Create using transfer learning with ResNet and the time were similar. Training using a complete convnet would take much long, unless the Lobe team has made some serious advances.
While the UI is quite nice (thanks to Mike Matas, I am sure), I don't see a strong advantage to using this on MacOs, when CreateML is available. CreateML doesn't have the simple interface of Lobe, but the UI is quite accessible and gives you access to additional classifies, like sound, text and tabular data. If you need ever more power, you can use TuriCreate if you want to stay in the Apple ecosystem.
The simplicity of the UI is a feature, but also a disadvantage when you start having more than a handful of labels and training images. I totally see how Lobe could be a nice intro into the world of labelling and classification.