Currently, machine learning is still very dependent on big data centers. Implementing artificial intelligence directly on edge devices, by contrast, brings advantages in terms of both data protection and efficiency. In general, however, two problems stand in the way: for one thing, battery-powered devices have a rather limited energy budget; having a powerful graphics card active in the background would quickly exceed it. This is especially true for devices and applications that are permanently in operation, such as a voice recognition system
The second problem is that machine learning requires huge data sets that simply don't fit on the available memory of a normal edge device. At present, only cloud providers can store datasets that are big enough for machine learning. However, this leads to privacy and security concerns: no one wants to send raw audio files, for instance, directly to the cloud of one of the major speech recognition providers.
SEC-Learn is designed to overcome these problems. To meet energy efficiency requirements, devices need to be able to handle data processing locally while making use of energy efficient dedicated neural network circuits. To this end, Fraunhofer IIS is developing a neuromorphic chip that is much more energy efficient than conventional chips. The other problem, i.e., how to take data that originates locally and pool it in a cloud while maintaining high security standards, is solved by federated learning. This means that no raw data has to be passed on, only the changes to the models.