What is neuromorphic about neuromorphic computing?

Abstract

(2025). The hardware has kept pace with the literature. Examples include IBM's TrueNorth integrated a million digital spiking neurons on a single die Merolla et al. (2014); Intel's Loihi processors lie at the core of Hala Point, a system exceeding a billion neurons Davies et al. (2018); SpiNNaker and BrainScaleS have matured into second-generation, wafer-scale and supercomputer-scale platforms Furber et al. (2014); Pehle et al. (2022); and memristive crossbar arrays promise in-memory analog computation Chua (1971) ;Strukov et al. (2008). Startups, national initiatives, and semiconductor roadmaps now routinely list "neuromorphic" among the post-Moore alternatives.Yet beneath this boom lies an awkward foundational question: what, exactly, is neuromorphic? Classical computing rests on a crisp abstraction, namely, Boolean logic, the Turing machine, the von Neumann architecture. This abstraction is entirely indifferent to whether it is realized in relays, vacuum tubes, or emerging fabrication technologies. Quantum computing likewise has a precise essence: state vectors in Hilbert space, unitary evolution, entanglement, and measurement, a definition sharp enough to support rigorous claims of quantum advantage Arute et al. (2019). Neuromorphic computing has no comparable statement of principle. "Brain-inspired" feels like a family resemblance, not a definition. The field itself has begun to sense this and a wave of recent editorials and perspective articles now asks what the essence of neuromorphic or neuromorphicism. Most pointed is Indiveri's editorial "Neuromorphic is dead. Long live neuromorphic", which argues that the label has drifted far from its founding meaning Indiveri (2025).History clarifies what the early pioneers of the field intended. Carver Mead coined "neuromorphic" electronic systems Mead (1990), in the late 1980s to describe circuits that exploit the physics of transistors operated in the subthreshold regime, where channel current depends exponentially on gate voltage in close analogy to ionic conductances in neuronal membranes. His 1990 manifesto argued that biology computes with 10,000-fold to billion-fold less energy than digital logic precisely because it uses device physics rather than fighting it Mead (1990). John Hopfield showed that computation could be a collective, emergent property of a dynamical system relaxing in an energy landscape Hopfield (1982), Sejnowski and Hinton introduced statistical mechanics into neural computation Ackley et al. (1985), contributions that were in retrospect worthy of a Nobel Prize in Physics Hopfield and Hinton (2024). Richard Feynman, who taught the famous course on the physics of computation alongside Mead and Hopfield, asked what physics itself can compute and where the thermodynamic limits of computation lie Feynman (1982Feynman ( , 1996)). What these pioneers of physics and computing intended was not the imitation of neurons but the grounding of computation in physical dynamics and energy. Physics of Computation is a field that flourished in the mid-1990s and has since morphed into "Unconventional Computing," an all-encompassing name and research field, one that certainly includes neuromorphic computing.This editorial addresses a misconception about computation in the current neuromorphic engineering discourse. A computing principle must abstract away from the physical substrate on which it is implemented, an idea introduced by the late David Marr Marr (1982), and subsequently addressed by Cassidy et.al. Cassidy et al. (2013) separating form and function. Statements such as "uses subthreshold CMOS," "uses memristors," or "uses neural processing units" are statements about implementation i.e. form, not essence, just as "uses transistors" never defined classical computing, and "uses superconducting junctions" does not define quantum computing. If neuromorphic computing is to be a paradigm rather than a product category, its defining principle must survive translation across substrates: silicon, oxides, photonics, or wetware.What, then, is the substrate-independent essence for neuromorphic computing ? The most defensible candidate is "thermodynamic". Nervous systems evolved under severe metabolic budgets: the human brain runs on roughly 20 watts, and cellular-level energy accounting shows that spiking, signaling, and synaptic transmission are all tuned to minimize metabolic cost Attwell and Laughlin (2001); Laughlin et al. (1998). Insect brains sharpen the point. A honeybee forages over kilometres, navigates by landmark and sky compass, communicates, and learns with about a million neurons, and a fruit fly flies, and adapts with roughly a tenth of that, on power budgets in the microwatt-to-milliwatt range. Direct measurements in fly photoreceptors show that the energy-information trade-off is enforced down to the level of single cells Abshire and Andreou (2000); Niven et al. (2007). The insect brain is therefore the gold standard for edge-AI embodying full sensing, autonomy, and lifelong learning within a milliwatt power-budget, a benchmark that no current engineered system, neuromorphic or otherwise, comes close to matching. Addressing function, neuromorphic computation is thus, at its core, computing with and against thermodynamics: representing problems as energy functions and entropy production, letting physical dynamics perform descent, relaxation, and sampling, and treating statistical mechanics as the native mathematics of the machine. In this sense neuromorphic computing dovetails with the emerging research programs of thermodynamic computing, which take fluctuations, dissipation, and stochastic dynamics as and, in particular, with the field of stochastic thermodynamics, where fluctuation theorems, uncertainty relations, and thermodynamic speed limits now make it possible to quantify the energy costs of computation in systems arbitrarily far from equilibrium, from digital circuits to brains Wolpert et al. (2024). Indeed, it is now known that any constraint, whether structural, architectural, or physical, produces entropy or dissipates energy. Therefore, the essence of neuromorphic computing could be "thermodynamic" mechanisms that optimize parallel processing under physical constraints of energy, delay, and structural variability Cassidy et al. (2013); Gangopadhyay and Chakrabartty (2021).To gain acceptance, such a thermodynamics-based framework would first need to explain the four principles that many in the field currently consider the fundamental signature and starting point of neuromorphic computing: (a) Event-based, asynchronous representation, where computation, sensing, encoding and communication is unified Boahen (2022); Gangopadhyay et al. (2020) and information is carried by the timing of sparse events rather than by clocked global state Purohit and Manohar (2022); Zhou et al. (2025); Srivatsav et al. (2023); (b) Distributed and massive parallelism, with memory and processing collocated Pouliquen et al. (1997); Andreou et al. (2026) within millions to billions of simple interconnected units Cauwenberghs (2013). (c) Analog processing, which extracts precision per joule from continuous representation that exploits device physics and locality Sarpeshkar (1998); and most importantly (d) Noise-induced transitions and fluctuations, in which thermal and device noise and possibly quantum fluctuations drive escape from local minima, probabilistic inference and stochastic computation Mendat et al. (2015), and exploration, making noise a computational resource rather than an error to be suppressed Buesing et al. (2011); Chen et al. (2025).Architecturally, the framework should then provide a scientific rationale for building machines based on multi-time-scale dynamics: fast membrane-like state variables, slower synaptic traces, and still slower homeostatic and structural processes, all coupled through local, continual learning rules and network oscillations. In such a machine there is no distinction between inference and learning. Both are the same physical dynamics viewed at different time scales, with no separate training phase, and no offline weight download. Consequently, everything observed and measured at the system level like polychronicity, or predictive and efficient coding are emergent phenomena. Synchronization across modular units, for instance, is not programmed but arises from coupling and possibly due to constraints on system entropy production, in the way oscillatory assemblies emerge across columns of cortex Hopfield (1982); Gray et al. (1989).Attention, likewise, is not a scheduled control signal but an emergent allocation policy, guided by the need to improve system-level energy efficiency by concentrating the metabolic budget (sparse activity, selective routing, selective plasticity) to where information matters most Lennie (2003).Practically, the thermodynamics-based framework and related design principles would be considered meaningful only if they scale. Furthermore, scaling must be defined not only in terms of the number of parameters, neurons, synapses, dendrites, or astrocytes, but also in terms of functional diversity. A dynamical system comprising a few hundred analog units and variables is a laboratory curiosity, whereas the brain's existence proof involves more than 80 billion neurons. Scaling from hundreds of units and variables to billions requires overcoming device mismatch, wiring constraints, thermal budgets and the absence of global clocks or global error signals, all of which is not a downstream engineering detail but a mainstream research problem Kudithipudi et al. (2025); Mehonic and Kenyon (2022). However, this is only one axis of scaling. A fruit-fly brain with about 50 million synapses can exhibit a remarkably wide diversity of functions, and yet the field of neuromorphic computing, and the traditional field of AI are engaged in a rat-race of packing more compute and more functional units, with some of the modern AI models incorporating more than a trillion synaptic weights. Unfortunately, cracks are already starting to show in the standard roadmap with respect to AI's exorbitant power and energy demands, and it can be envisioned that the field will eventually revisit, introspect and incorporate the neurobiological "form" that embodies modularity, 2.5D/3D integration Marwick and Andreou (2006), and hierarchy, in the context of thermodynamic principles Wolpert et al. (2024).Finally, the hope is that the framework will impose upon the neuromorphic paradigm the same discipline that quantum computing imposed on itself: a clearly articulated neuromorphic advantage. Quantum researchers are expected to identify problems where quantum mechanics provably or plausibly beats the