NeuralTalk takes in and interprets its surroundings on a street in Amsterdam.
Source: Popular Mechanics
Bytes on Neural Networks: In contrast to conventional computing done through a serial computer (which relies on a central processor that accesses and stores data in memory locations and executes instructions programmed by a person), neural networks are algorithms that react to the information presented to them in un-programmed ways. Neural networks can react in un-programmed ways (they can learn) because of how they are structured. A neural network's set of "input units" can communicate with each of its "output units" simultaneously. An "input unit" receives data from the outside world (an image, a typical data table, unstructured sources of data scraped from the internet, etc.) and interacts with the neural network's "output units", which react to the information processed by the input units. Because neural networks can learn dynamically and construct knowledge through approximated, imprecise data sets after iterative data feeds, one of the more popular use cases for these networks has been image recognition.