Classifying images

Topics: AI (Deep Learning), Document Classification, Image Search, Jeff Dean

The Google patent describes a system for classifying images using neural networks and high-dimensional representations. The system works by converting input images into high-dimensional numeric representations and comparing them to pre-associated category labels in the same high-dimensional space. The key innovation is using a transformation layer after the core neural network layers to generate these representations, which allows accurate classification even for categories not seen during training. The system matches input image representations to the closest category label representation based on similarity measures like cosine similarity. Benefits include improved zero-shot predictions and semantically meaningful incorrect predictions.