Deep learning algorithms are instrumental in many artificial intelligence use cases, including image recognition. A car’s ability to recognize a pedestrian, for example, depends on deep learning. And how well deep learning can perform depends on the availability of a large amount of “tagged data.” For pedestrian recognition, the data would be images of pedestrians that have been identified, or “tagged” as such. The cost of learning, however, is often prohibitive for system manufacturers. CEA-List Carnot institute recently developed a semi-supervised learning process capable of improving the system’s performance without the need for more tagged images.
Inspired by human learning, semi-supervised learning can take note of similarities between objects to extrapolate new knowledge. The challenge here was to give neural networks the ability to learn from a large amount of data with just a small portion tagged. The data was organized “geometrically” to make the extrapolation of knowledge easier. Specifically, this organization allows the algorithm to find a similar tagged object for each non-tagged object encountered.
Using a database with just 25% tagged data, the neural network was 89% accurate, earning the method a respectable position among the different semi-supervised learning methods currently available.
Next, the researchers plan to improve the ergonomics and computational efficiency of the technique to facilitate its application to industrial use cases.
Read article at http://www.cea-tech.fr/