Core ML 3 runs entirely on device, is hardware accelerated (neural networks), and can update models locally as well.
Domains updates (reminder: all of this runs on device now):
Vision Updates
Image classification
Image Similarity
Face Capture Quality
Human Detection
Object Saliency: vision can figure out the main subject of a given image and let you know where it is in the image
Attention Saliency
Text Recognition
Animal Detection
Document Camera (Document recognition and auto OCR)
Improved Object Tracker
Improved Face Landmarks
Natural Language Updates
Sentiment Analysis: tells if a given phase is positive/negative etc.
Word Embeddings: tells words similarity (topic-wise, e.g. cloudy and Thunderstorm)
Text Catalogs
Transfer Learning
Speech and Sound Updates
Speech Recognition
Voice Analysis: beside telling you what has been spoken, voice analysis provides also how it is spoken (normal voice and more)
We can combine different CoreML Domains seamlessly, for example:
categorize images and assign tags to them
use speech recognition to catalog text and find the images that match that categorization using word embeddings.
Improved compatibility with other ML creation tools (TensorFlow and others).
Model Fine-Tuning on device
It is now possible to create personalized model per each user, privacy, no server costs.
This can be done in the background (see
BackgroundTasksframework).
