Neuromorphic Computing Gets a Common Language for Comparing Brain-Inspired Machines
Neuromorphic engineering has entered a period of rapid expansion, with researchers building computing systems that imitate different features of the nervous system. Some devices reproduce the way synapses combine signals, while others communicate through brief electrical pulses, process information within dendrites or adapt their behaviour through plasticity. Yet this diversity has created a serious problem: comparing one neuromorphic platform with another is often like comparing a bicycle, a jet engine and a biological muscle using the same ruler. A new Review in Nature Reviews Electrical Engineering proposes a unified way to evaluate these technologies by asking a fundamental question: how effectively does a device emulate a specific biological computation?
The work, by U. Bruno, M. Farronato, A. Salleo and colleagues, addresses a challenge that has become increasingly urgent as neuromorphic systems move from laboratory demonstrations toward practical applications. Conventional computers are usually compared using familiar measures such as processing speed, memory capacity and energy consumption. Neuromorphic systems, however, do not all perform computation in the same way. They may calculate with continuous signals, discrete spikes, changing conductances or networks of adaptive elements. A single metric can therefore reward one architecture while overlooking the feature that makes another biologically meaningful or technologically valuable.
The Review organizes this complicated landscape around different levels of computation found in living nervous systems. At the smallest scale, a synapse receives and transforms signals from neighbouring neurons. At the level of a neuron, dendrites collect inputs, the cell body integrates them and the neuron produces an output spike when its internal state reaches a threshold. At larger scales, networks learn, adapt and coordinate activity through changing connections. Neuromorphic devices can reproduce one or more of these operations, but they do not necessarily emulate the entire biological system. The authors argue that evaluation should focus on the function actually reproduced rather than on whether a device is simply labelled “brain-inspired.”
That distinction could reshape how researchers describe progress. A material that changes its electrical resistance after receiving a stimulus may emulate a form of synaptic plasticity, even if it does not reproduce every molecular process in a biological synapse. A circuit that integrates incoming pulses and generates thresholded outputs may emulate a neuron-like computation without matching the full complexity of a living cell. By separating biological function from physical implementation, the proposed framework can accommodate many technologies, including electronic circuits, emerging materials and hybrid devices. It also makes it possible to compare systems that use entirely different mechanisms but perform related computational tasks.
One of the central ideas is biological emulation efficacy: the degree to which a neuromorphic device captures the computational behaviour that gives a biological structure its usefulness. This concept goes beyond visual resemblance or the use of neuroscience-inspired terminology. A system may contain artificial neurons and synapses yet offer little biological relevance if its components cannot reproduce the timing, adaptation or signal-processing behaviour that matters for the target function. Conversely, a relatively simple device may be highly effective if it faithfully performs a specific operation with low energy use, compact hardware or rapid response.
The Review also highlights why energy efficiency remains one of neuromorphic engineering’s most attractive promises. In the brain, computation takes place through vast networks of interconnected cells while consuming remarkably little power compared with many conventional high-performance systems. Neuromorphic architectures seek to approach this efficiency by processing information where it is stored, reducing data movement and using event-driven communication. In a spiking system, for example, a neuron may remain quiet until a meaningful event occurs, rather than being updated continuously. This can reduce unnecessary operations, particularly for sensory tasks involving sparse or rapidly changing data.
Spiking communication is only one part of the picture. Dendritic processing can allow a single artificial neuron to perform more complex operations before producing an output, potentially reducing the number of separate processing elements required. Synaptic computation can combine information locally, while neuroplasticity can enable systems to adjust their responses in reaction to experience. These features may be important for applications in which data arrive continuously and unpredictably, including robotics, autonomous sensing and intelligent interfaces. The Review presents such capabilities as distinct dimensions that should be measured according to the task a system is intended to perform.
A unified comparison also requires attention to practical engineering constraints. An architecture that closely mimics biological behaviour may be difficult to scale, manufacture or program. Another system may be less biologically detailed but offer higher reliability, greater integration density or a clearer path to commercial deployment. The authors therefore frame neuromorphic performance through several complementary priorities, including bioplausibility, energy efficiency and scalability. Rather than searching for a single universal winner, the proposed approach is designed to reveal which architecture is strongest for a particular application and which compromises it makes.
This perspective could prove especially valuable as neuromorphic research expands beyond isolated demonstrations. Future systems may combine conventional digital processors with analogue components, memory devices, spiking circuits and adaptive materials. Without common definitions and figures of merit, claims about efficiency or intelligence can be difficult to interpret across platforms. A device optimized for learning may not be optimized for speed; a system designed for biological realism may consume more resources than one built for a narrow industrial task. Evaluating each platform through the biological function it emulates, alongside its technical performance, could make comparisons more transparent and guide more targeted development.
The Review ultimately presents neuromorphic engineering as a field that needs both inspiration and measurement. Neuroscience can suggest powerful computational principles, but translating those principles into useful machines requires clear criteria for judging what has actually been reproduced. By establishing a shared framework for examining synaptic computation, spiking communication, dendritic processing and plasticity, the authors aim to help researchers identify where each technology has a genuine advantage. The result could be a more disciplined race toward low-power computing—one in which the most successful machines are not necessarily those that look most like the brain, but those that emulate the right biological functions for the job.
Subject of Research: Neuromorphic devices and architectures, with a focus on evaluating their effectiveness in emulating biological computation.
Article Title: Figures of merit for neuromorphic devices through biological emulation efficacy
Article References: Bruno, U., Farronato, M., Salleo, A. et al. “Figures of merit for neuromorphic devices through biological emulation efficacy.” Nature Reviews Electrical Engineering (2026). https://doi.org/10.1038/s44287-026-00321-7
Image Credits: AI Generated
DOI: 10.1038/s44287-026-00321-7
Keywords: Neuromorphic engineering, neuromorphic devices, brain-inspired computing, biological emulation, synaptic computation, spiking communication, dendritic processing, neuroplasticity, energy-efficient computing, scalability
