Google’s multitasking neural net can juggle eight things at once




Deep-learning systems tend to be one-trick wonders: they’re great at the task they’ve been trained to do, but pretty awful at everything else. Now a new neural network from Google suggests that AI can be taught to multitask after all.
Most deep-learning systems are built to solve specific problems, such as recognising animals in photos from the Serengeti or translating between languages. But if you take, for instance, an image-recognition algorithm and then retrain it to do a completely different task, such as recognising speech, it usually becomes worse at its original job.
Humans don’t have that issue. We naturally use our knowledge of one problem to solve new tasks and don’t usually forget how to use a skill when we start learning another. Google’s neural network takes a tiny step in this direction, by simultaneously learning to solve a range of different problems without specialising in any one area.
The neural network from Google Brain – one of the search giant’s deep-learning teams – learned how to perform eight tasks, including image and speech recognition, translation and sentence analysis. The system, called MultiModel, is made up of a central neural network surrounded by subnetworks that specialise in specific tasks relating to audio, images or text.
Advertisement: Replay Ad

Consistent performance

Although MultiModel did not break any records for the tasks it attempted, its performance was consistently high across the board. With an accuracy score of 86 per cent, its image-recognition abilities were only around 9 per cent worse than the best specialised algorithms – matching the abilities of the best algorithms in use five years ago.
The system also showed other benefits. Deep-learning systems usually need to be trained on large amounts of data to perform a task well. But MultiModel seems to have come up with a neat way of sidestepping that, by learning from data relating to a completely different task.
The network’s ability to parse the grammar of sentences, for example, improved when it was trained on a database of images, even though that database had nothing to do with sentence-parsing.
Sebastian Ruder at the Insight Centre for Data Analytics in Dublin, Ireland, is impressed with Google’s approach. If a neural network can use its knowledge of one task to help it solve a completely different problem, it could get better at those that are hard to learn because of a lack useful data. “It takes us closer on the way to artificial general intelligence,” he says.
Google has released the MultiModel code as part of its TensorFlow open-source project, giving other engineers a chance to experiment with the neural network and put it to the test. The network’s complexity, however, might make it difficult for researchers to work out the reason behind its multitasking skills, says Ruder.
Journal reference: arxiv.org/abs/1706.05137

No comments:

Powered by Blogger.