As document processing automation professionals, we want to look at a concept that you may have heard talked about but are not sure what it means: NPUs What are they? What are their applications?

The first thing to make clear is that, when we speak about NPUs, Neural Processing Units, we are not referring to neuromorphic computing. The Artificial Intelligence (AI) acceleration chips that are used, for example, for optimising management of resources in smartphones. These chips vary depending on the environment in which they are used. For example, they will not be the same in graphics, the audiovisual field or in neural networks. In the first case, there are GPUs – graphic processing units. In the second, DSPs – digital signal processors. But those that really interest us at Serimag, those in which we are competent, are the chips found in neural networks; the NPUs (Neural Processing Units)

Neural networks are used to attempt to emulate the synaptic connections of the human brain. NPUs are already present in devices of brands such as Huawei and Apple. AI is not only used to manage the resources of the mobile, but the device learns from the user by using neural networks. It also makes use of machine learning. Based on this philosophy, there are an increasing number of smartphones capable of learning what the users want and adapting themselves to their lifestyles.

As you see, NPUs are faster and more efficient than traditional chips, which is very important in mobiles. It must be borne in mind that although they are increasingly common in smartphones, not all Operating Systems accept them. The Android 8.1 does. A further example of the progress of machine learning processes is the automatic learning of systems. These are based on algorithms that review the data in order to predict future behaviour.

NPUs also have important applications in one of our star products: SMAAT. The product uses these applications for machine learning and deep learning processes for capturing documents through smartphones, a principal functionality of this technology.