But what about technology? Researchers have known for years that there is a limit to how much data can be collected and processed
Now a new study has suggested that these limits might be artificial, and that a whole new generation of artificial neural networks could eventually overcome them.
Researchers from MIT and the Massachusetts Institute of Technology (MIT) and IBM developed artificial neural networks that could solve problems that are too complex for traditional machine learning techniques to solve.
“Our approach takes a completely different approach to solving these problems,” says MIT computer scientist and lead author Arvind Narayanan. “Rather than focusing on specific tasks, we developed a system that learns how to solve any problem. This way, we can keep increasing the level of generality of the networks as the problem becomes more complex.”
These artificial neural networks (ANNs) learn how to do tasks by taking in information about how they should behave, and then adapting their own software to achieve the goal.
These ANNs were able to outperform conventional machine learning algorithms, such as neural networks, in solving a variety of difficult problems, such as recognizing handwritten digits, sorting images into categories, and categorizing languages.
“The study shows that the algorithms we have created can learn complex problems that would have previously required sophisticated and specialized approaches,” says Narayanan. “This is important because if you want to apply these algorithms to other domains, such as robotics or medical diagnosis, you need to be able to tackle those problems at the same level.”
An exciting future
The researchers say their new ANNs could help solve some of the biggest problems facing AI, such as improving self-driving cars and recognizing facial expressions. But Narayanan says they could also be used to help tackle more common problems, such as understanding how everyday objects are used in our everyday lives.
“The great thing about ANNs is that they can be adapted to work with different data sets,” says Narayanan. “This means that you could use them to identify new uses for everyday objects and objects in your environment.”
Narayanan says the paper, “Understanding machine learning with deep neural networks,” is a very exciting advance for the field.
“The approach we have taken has the potential to dramatically improve the efficiency of many machine learning algorithms,” he says. “We are excited about the direction this research is going in and look forward to seeing how it will evolve.”
Explore further: MIT AI Lab’s algorithms boost machine learning
More information: ‘Understanding machine learning with deep neural networks’ is available at arXiv:1605.00208 [cs.NA] or at dx.doi.org/10.1088/1741-2560/1605/2/4