The challenge of building, within a few decades, a computer chip on the scale of a patch of biological cortex is a race involving many labs in academics and industry around the world.

The basic assumption is that, in order to build machines that imitate the cortex, the intuitive way to go is capture in a chip the architecture and functional principles of cerebral cortex. Building a chip that emulates the cortex needs to solve several challenging problems. For example, how can you pack millions of processing elements and billions of synapses into a small enough chip and be able to perform computations at a speed compatible with human thought. All this must be done without consuming a lot of power. Easy, right?

Fortunately, biology comes to the rescue. One of the characteristics of the cerebral cortex is that neurons rely on discrete events, called spikes, which mediate most of the communication between neurons. There are some known exceptions to this rule, including neurons that communicate across gap junctions (e.g., in the retina), direct electrical coupling between neural membranes, and diffuse chemical transmission (e.g., acetylcholine). These exceptional cases enable neurons to influence the computations carried out by a large numbers of receiving cells. Nevertheless, the overwhelming majority of conversations going on between cortical neurons adopt the “spiking” dialect.

Surprisingly enough, most of the published research that attempts to model various aspects of cortical processing, from models of visual perception to executive and motor functions, have ignored the fact that neurons do communicate with spikes. These models assume that a neuron’s firing rate is the output of the neuron or the “dialect” that artificial neurons use to talk to each other within simulated cortical networks. This assumption bears some important implications, namely that the timing of single spiking events does not matter because the artificial neurons should only be concerned with the projecting neurons’ firing rate, rather than being influenced by discrete spiking events.

It turns out that there might be deeper reasons why the brain has adopted spikes as means of neural communication over the course of evolution. This strategy has even deeper implications for solving key issues faced by emerging neuromorphic technology: the metabolic cost of broadcasting spikes vs. analog values (e.g., firing rates) among neurons in a network. An encoding/decoding scheme that relies on discrete events, and their order of occurrence, rather than on broadcasting and estimating firing rates, is a more parsimonious, efficient, and faster neural communication protocol that might have been favored by evolution. Neuromorphic technology will most likely follow the trend in the quest to build smaller, faster and less power-hungry chips that emulate cortical networks. This also implies that the modeling community should consider building more models that embody spiking dynamics as basic building blocks of cortical networks.