Biologically plausible artificial networks of neurons take a number of different forms in current research. A researcher's choice of approach relates to which of the biological factors they wish to mimic most strongly. Convolutional neural networks (CNN), for example, are based on the visual cortex. Spiking neural networks (SNN), based on the brain’s asynchronous processing, allow each neuron to fire independently of the others and offer a greater efficiency than synchronous networks. SNN is the architecture used by the SpiNNAker project at the University of Manchester,⁷ as well as by IBM (the TrueNorth chip) and Intel (the Loihi chip). There is also some interest in using photonics, rather than electronics, in neuromorphic networks.⁸,⁹ Different hardware configurations seem to provide different capabilities. Intel’s Loihi is highly configurable for specialised applications,¹⁰ for example, and IBM’s TrueNorth is particularly suited to high-speed and low-energy image processing and classification tasks.¹¹

In addition to these mainstream efforts, there are several alternative technological approaches. Some emerging architectures involve memristors, for example, which are simple transistor-like components that have variable resistance and the ability to store multiple memory states. Several dozen AI start-ups are also developing different architectures. It is not yet clear how soon — or indeed whether — we will begin to see convergence between these efforts. To mimic the brain, all will need to find ways to efficiently integrate learning and memory in hardware elements.