Most deep-learning Software is built to solve specific problems, such as identifying animals in photos from the Serengeti or translating among languages. But if you take, for example, an image-recognition algorithm and later retrain it to do a completely different task, such as understanding speech, it usually becomes worse at its initial job.
Humans don’t have this issue. We naturally use our experience of one problem to solve new tasks and don’t normally forget how to use a skill if we start learning another. Google’s neural network takes a small step in this direction, by simultaneously learning to answer a range of different problems without training in any one area.
The neural network interface from Google Brain one of the search giant’s deep-learning team working on this Platform learned how to perform eight tasks, covering image and speech recognition, translation and sentence analysis. The operation called MultiModel is made up of a central neural network enclosed by subnetworks that train in specific tasks relating to audio, images or text.
Consistent performance
Despite MultiModel did not break any records for the tasks it tried, its performance was consistently large across the board. With a precision score of 88.6 percent, its image-recognition abilities remained only around 9 per cent worse than the best-specialized algorithms matching the abilities of the best algorithms in use five years ago.
The operation also showed other benefits. Deep-learning systems usually must to be trained on large amounts of data to perform a job well. But MultiModel seems to have come up with a smart way of sidestepping that, by reading from data relating to a completely different task.
The network’s ability to parse the grammar of sentences, for instance, improved when it was exercised on a database of images, even though that database had nothing to do with sentence-parsing.
Sebastian Ruder at the Insight Hub for Data Analytics in Dublin, Ireland, is impressed with Google’s way. If a neural network can use its data of one task to help it solve a completely different problem, it could get bigger at those that are hard to learn since of a lack useful data. “It takes us closer on the way to artificial general intelligence,” he says.
Take your time to comment on this article.
