Teaching the Machine, Not You
Newly released app Lobe takes the difficulties out of machine learning with a three-step process and simple interface.
If you have ever had a great idea for a machine learning application, then went cross eyed and gave up after reading a tutorial about the mathematical underpinnings of a learning algorithm, then you may want to check out the recently launched Lobe application and dust off that old idea.
Lobe is an application that is designed to make image classification with machine learning easier for the uninitiated — no coding or previous experience with machine learning is required. The software walks a user through three steps: collecting and labeling data, training a model, and “playing,” which is an interactive form of model optimization.
Collecting data is accomplished by either using a webcam to capture new images of each class you wish to represent, or by loading existing images stored on your computer. Labels can be attached in either case. Training is an automated process that requires no selection of model type or hyperparameter values. Lobe takes advantage of transfer learning, in which a pre-trained model is used as a starting point, to allow training to happen on the local machine, and with a relatively modest set of training data. The play mode allows for real-time inference from the model that was just trained. By clicking green or red icons, the user can tell Lobe if the prediction was a good one. If not, a new model is trained in the background with the benefit of that additional information.
Finalized models can be exported for use in other machine learning frameworks, such as TensorFlow. These models can then be integrated into custom applications on a desktop or embedded device, such as a Raspberry Pi.
Lobe certainly makes creating a machine learning algorithm very easy by hiding the details. A cost of that convenience is that the user will be largely unaware of the process that went into building the model, and when they run into a situation where hyperparameters need to be tweaked, or a specific model needs to be used, they will have gained no understanding of how to accomplish these goals from their time using Lobe. That is an important consideration with any convenience software such as Lobe — do you want only the machine to learn, or do you want to learn something as well?
If you would like to give Lobe a try, you can get a copy here.