Sensing, Encoding, and Memory in a Self-Powered Analog Neuromorphic System

Kim, S. J. et al. Self-powered analogue neuromorphic system for multimodal sensing, encoding and learning with diffusive and drift memristors. Nat. Sens. 1–10 (2026) doi:10.1038/s44460-026-00067-7.

In conventional electronic systems, environmental signals are usually first acquired by sensors, then converted through analog-to-digital conversion, sent to a digital processor, and finally stored in memory. This pipeline is clear and highly successful. But it also rests on an implicit premise: sensing, encoding, computation, and memory are separate modules. The sensor is merely the entry point; computation happens in the processor.

The article by Kim et al. proposes a different path. The authors construct a fully analog, self-powered neuromorphic system that integrates multimodal sensing, spike encoding, unsupervised learning, and memory storage within the same hardware platform. The system consists of light sensors, pressure sensors, diffusive memristors, drift memristors, and a small number of passive components. It requires no ADC, no digital controller, and no external processing unit. Environmental stimuli not only provide information, but also directly supply the energy that drives the system.

What is most notable about this work is not whether it “resembles the brain,” but that it concretely demonstrates a more fundamental question: how information becomes internally usable structure through a physical interface. In previous blog posts, I discussed that information is not an abstract entity attached to symbols, but a physical correlation established through interaction between systems. What a system can access is not determined simply by what exists in the world, but by which external differences its interface allows to leave distinguishable traces in its internal states.

This article provides a hardware-level example of precisely this idea. External variables such as light intensity, pressure, and acoustic delay are not first abstracted into numbers and then interpreted by an algorithm. Instead, through the physical dynamics of sensors and memristors, they are directly transformed into spike timing, conductance changes, and long-term weights. In other words, the system does not first “acquire data” and then “perform computation”; its sensing interface already participates in encoding and learning.

The core devices in the article are of two types. The first is the diffusive memristor. It exhibits volatile threshold-switching behavior: when the input voltage is sufficiently high, silver ions form a conductive filament, causing the device to switch on abruptly; the filament then spontaneously dissolves through diffusion. Because this process depends on ion migration time, stronger input shortens the time required to reach the threshold. The diffusive memristor can therefore naturally transform analog input intensity into spike timing.

The second is the drift memristor. It has non-volatile conductance modulation capability: under an applied voltage, its conductance can be gradually changed and then retained. It is therefore suitable as a synaptic weight storage unit. The basic design of the article is straightforward: the diffusive memristor converts input into temporal structure, while the drift memristor stores these temporal relations as long-term weights. The former corresponds to encoding; the latter corresponds to memory.

The authors first validate the analog-to-spike conversion capability of diffusive memristors using two modalities: light and pressure. In the light-stimulation experiment, a photovoltaic sensor is directly connected to a diffusive memristor. The stronger the light intensity, the higher the voltage generated by the sensor, the faster the memristor switches on, and the earlier the spike occurs. In the pressure-stimulation experiment, a PVDF piezoelectric film converts mechanical pressure into voltage signals. Greater pressure likewise produces shorter spike timing. The entire process does not rely on amplifiers, ADCs, or external power; the voltage generated by the sensor itself is sufficient to drive memristor switching.

The significance of this step is not merely that the authors “implemented a low-power encoder.” More precisely, they transformed differences in external stimulus intensity into temporal differences that can be compared within the system. Light intensity and pressure originally belong to different physical modalities, but once both are converted into spike timing, the system obtains a common dimension for comparison. This is exactly the role of an interface: it does not infer what the world is, but transforms certain differences in the world into internal variables that are distinguishable, transmissible, and learnable.

The article then demonstrates two forms of synaptic plasticity. The first is homosynaptic plasticity, namely classical STDP. The authors construct a circuit using one diffusive memristor and one drift memristor, allowing the relative timing between pre-synaptic and post-synaptic spikes to determine the conductance change of the drift memristor. Different spike orders and time intervals lead to different directions and magnitudes of potentiation or depression. Here, learning is not computed by a digital algorithm, but directly implemented by the device’s physical response to spike timing.

The second is heterosynaptic plasticity. The authors use two diffusive memristors and one drift memristor so that two input pathways generate spikes separately. If the two input amplitudes are similar, the spike times converted from them are also close, and their overlap produces stronger modulation of the drift memristor. If the two inputs differ greatly, the spike times are separated, and the resulting weight change is weaker. Experimentally, the conductance change of the drift memristor shows a Gaussian-shaped dependence on the input-amplitude difference, resembling input-timing-dependent plasticity.

This step is crucial. The system does not separately record the strengths of the two sensors. Instead, it records their relationship at the hardware level. In other words, what is learned is not a single input, but the temporal correlation between multimodal inputs. The diffusive memristors convert amplitude differences into timing differences; the drift memristor converts timing differences into storable weight differences. External correlations thus become internally memorable structures.

This also connects to the point I made in my previous blog post on “semantics.” Semantics is not internal complexity itself, but a stable and usable correspondence between internal structure and external structure. An internal difference has usable meaning only when it stably tracks an external difference and can potentially influence subsequent behavior or task outcomes. In this article, the correspondence between light and pressure is not yet full behavioral semantics, but it already provides the first half of semantic formation: external structure induces distinguishable internal states, and these states can be stored as weights.

The article further integrates this mechanism into a self-powered sensor fusion unit. This unit contains a light sensor, a pressure sensor, two diffusive memristors, one drift memristor, and passive components such as resistors and capacitors, all integrated onto a small PCB. Light and pressure inputs separately drive the two diffusive memristors, generating two spike times. The relative relation between these spikes then modulates the conductance of the drift memristor, thereby writing multimodal correlations into a non-volatile state. Under fixed light input and varying pressure input, the authors observe that when the two inputs are better matched, the conductance change of the drift memristor is larger; when the two inputs differ more, the conductance change is weaker.

This is the concrete meaning of “sensing, encoding, and learning” in the title of the article. Sensing is not an isolated input step; encoding is not a digital format conversion; learning is not an optimization process performed by an external processor. The three are compressed into a single physical chain: environmental stimuli generate voltage, voltage changes ion migration time, temporal relations change conductance states, and conductance states store multimodal relations.

The authors also use two application-level simulations to illustrate the potential uses of this mechanism. The first is lightning-strike localization. Lightning produces both light and sound: light arrives almost instantaneously, whereas sound arrives with a delay due to its slower propagation speed. If multiple self-powered sensor fusion units are distributed in space, they can use the light-sound time difference to change internal weights, and these weights can then be used to reconstruct the location of lightning strikes. The second application is human activity recognition. The authors construct a sensor fusion layer using EMG and insole pressure data. The results show that adding the fusion layer accelerates training and improves test accuracy from 58% to 79%.

These applications are not the core hardware validation of the article, but they clarify the type of scenario the authors are targeting: long-term, low-power, near-sensor, edge-side information processing. In such scenarios, the most expensive part is often not the final classification, but the continuous acquisition, conversion, transmission, and storage of large amounts of raw signals. If part of the encoding and fusion can be directly performed near the sensors by physical devices, the system no longer needs to move every analog detail into a digital processor.

This is closely related to my previous discussion on the decoupling between energy cost and information complexity. The energy cost of an interface mainly depends on how many distinguishable states it maintains, how many variables it tracks, and with what precision it sustains coupling. It does not directly depend on the complexity of the input structure itself. Once an interface exists, complex temporal patterns, correlations, and higher-order structures can enter the system through environmental dynamics, rather than having to be generated internally from scratch.

The system proposed by Kim et al. embodies this point in hardware. Light fields, pressure changes, light-sound delays, and cross-modal correlations in human movement already exist in the environment. The system does not use large-scale digital computation to reconstruct these structures. Instead, it allows these structures to directly change internal states through the physical coupling between sensors and memristors. Energy consumption is mainly used to maintain the interface and device response, rather than to enumerate, reconstruct, or search over all possible multimodal relations inside a processor.

Therefore, this article can be understood as a “small-scale interface learning system.” It does not contain a complete intelligent agent, complex behavior, or a genuine semantic closed loop. But it reveals an important intermediate layer: between the environment and digital computation, there can exist a physical interface composed of materials, sensors, and temporal dynamics. This interface can transform external structure into internally storable differences. It is not intelligence in full, but it is one necessary component of an intelligent system.

If we follow Shannon and Weaver’s three levels of communication, this article mainly addresses the boundary between Level A and Level B. Traditional communication theory concerns the reliable transmission of symbols, whereas intelligent systems must also confront the problems of meaning and use. In my previous blog on communication, I argued that Level B and Level C cannot be solved by signal transmission alone, because meaning and behavior depend on how a system interacts with its environment. This article does not yet solve Level C at the behavioral level, but it shows that the precondition for Level B can occur at the hardware level: signals are not passively transmitted, but organized into associable and storable structures as they enter the system.

Thus, the most important insight of this article is not that “memristors can simulate synapses,” but that “the sensing interface can take over part of the work of information organization.” Traditional computing architectures usually treat the interface as the input end and computation as something that happens inside the processor. Here, however, the interface itself already performs filtering, temporalization, association, and storage. It does not simply send the world into the system; at the very moment the world enters the system, it determines which differences can become learnable structures.

From my perspective, the significance of this article is not that it proves a memristive system already possesses intelligence, nor that it has realized the computational principles of the biological brain. Rather, it provides a clear physical example showing that the sensing interface is not merely a channel through which information enters the system. It can also be the place where information is temporalized, associated, and memorized. External stimuli enter the system through sensors, but they are not first abstracted into numbers and then computed by an internal processor. They directly alter the dynamics of the device, transforming external differences into internally distinguishable and storable states.

This makes one sentence in the article especially important:

Environmental stimuli serve not only as the information source but also as the energy source, allowing the system to autonomously sense, encode and learn from multimodal inputs in real time.

This sentence can be read as the central idea of the whole article. Environmental stimuli are not only sources of information, but also sources of energy. The system can autonomously sense, encode, and learn not because it contains a powerful central processor, but because the external world itself drives internal state changes through a physical interface. Information structure and energy input are not completely separated: the same external stimulus both carries distinguishable structure and provides the physical driving force that allows this structure to be written into the system.

This is also one key reason why I believe the brain is so energy-efficient. The brain does not face an unstructured data stream and then construct the world from scratch through internal computation. Instead, the world itself already contains rich spatiotemporal structure: light fields, sounds, pressure, motion, chemical gradients, and bodily feedback continuously drive the nervous system. Receptors, neurons, synapses, and the body jointly form multilayered interfaces that transform these external structures into spikes, phases, correlations, and state transitions within the nervous system. The brain spends energy maintaining the distinguishability and plasticity of these interfaces, but it does not need to pay an equivalent internal computational cost for the complexity of the external structure itself.

From this perspective, the efficiency of biological intelligence is not merely a matter of “low-power computation.” It is a deeper interface problem: how can a system allow environmental structure to directly participate in the formation of its own internal states? Kim et al.’s work provides a small-scale but highly intuitive hardware example. It shows that when environmental stimuli serve simultaneously as information sources and energy sources, sensing, encoding, and learning no longer need to be separated into distinct steps. They can become different aspects of the same physical interaction process.

In other words, intelligent systems do not process the world from outside the world. They are driven, constrained, and shaped through their coupling with it. A truly energy-efficient system may not be one that compresses its internal algorithms to the extreme, but one that designs the right interface, allowing external structure to enter internal dynamics at minimal cost.