Google is preparing a November release of its DeepMind unit's StreetLearn dataset for training machine learning models to navigate cities without a map.
StreetLearn covers multiple regions within London, Paris, and New York, composed of cropped panoramic images of street scenes that each forms a node in a larger network or graph of images.
Says DeepMind's Raia Hadsell, "Initially when the agent first starts learning, it's only going to be given targets that are nearby in its own neighborhood. Gradually those targets get further and further away until they're covering the entire city.
StreetLearn contains three neural networks, including a convolutional neural network to manage image recognition and input that data to two Long Short Term Memory (LSTM) networks. One LSTM selects the action the agent should take next based on its current reward state, while the other must memorize the local environment and gain representations of both the agent's current position and the location of its destination.
View Full Article
Abstracts Copyright © 2018 Information Inc., Bethesda, Maryland, USA