Credit: Midjourney
The brain does not wait for instructions. It senses the world, finds patterns, and learns from them in real time, without a power cord, a processor, or a data center. Building hardware that works the same way has been one of the hardest unsolved problems in computing. Most attempts have required bulky digital circuits, external processors, and a steady power supply. The result is systems that are too large, too slow, and too energy-hungry to work in remote or miniaturized settings.
A team at the University of Southern California has built something that works the way the brain does.
In a study published as the cover story of the new issue in Nature Sensors, researchers from the Ming Hsieh Department of Electrical and Computer Engineering at the USC Viterbi School of Engineering and the USC Stevens School of Computing and AI report a fully analog, self-powered neuromorphic system that can sense the physical world, learn from it, and make decisions, all on a circuit board the size of a coin, powered by nothing external. The work was supervised by Professor J. Joshua Yang, with postdoctoral researcher Seung Ju Kim as first author.
The implications reach far beyond the lab. A device that needs no battery and no connection to a network could be scattered across wildfire-prone terrain to detect lightning strikes. It could be embedded in smart glasses that process the world around them without ever connecting to a phone. It could be sent into space or dropped into the deep ocean and left to learn on its own, in places where recharging is impossible and sending data home takes too long to be useful.
“The signal is not only the signal to be processed,” Yang said. “It is also the source of energy to power the system.”
No battery, no computer, no conversion
The key to making the system self-powered is radical simplicity. The world is analog. Light, pressure, heat, sound — these signals arrive as continuous physical phenomena, not as ones and zeros. Every conventional computing system converts them into digital form before doing anything with them, and that conversion is expensive. It requires complex circuitry, takes time, and consumes power.
“The raw data is analog and huge at the beginning,” Yang said. “If you have to convert it to digital, then you need another complex circuit and a lot of energy to do that. And then you need to store it. Storage also takes a lot of time and energy and space.”
The system is built around two types of nanoscale devices called memristors, one that behaves like a neuron and one that behaves like a synapse. The neuron-like device fires a signal whose timing depends on the strength of the input — a brighter light or a harder press makes it fire faster. The synapse-like device holds onto what it learned, updating its state when signals arrive and retaining that state indefinitely once they stop. By wiring these two types of devices together with a handful of resistors and capacitors, the team built a circuit that can take signals from different sensors, find the relationship between them, and hold onto what it learned, without any software, without any digital processing, and without any external power.
Learning the way the brain learns
What makes the system unusual is not just what it does, but how it learns. Most AI systems today are trained using supervised learning: a model is shown millions of labeled examples until it gets the right answers. That process requires enormous computing resources and does not happen in real time. Yang’s system uses a different approach, one he calls natural intelligence rather than artificial intelligence.
“Our brain network learns more like unsupervised learning,” Yang said. “If you show a little kid a desk, a chair, a horse, a sheep, a pig, and you don’t tell them anything, very quickly they can classify which belong to animals and which to something else, even though they all have four legs. That is unsupervised learning. You learn from the data and understand the internal connection.”
The system works like memory does in the brain: experiences that repeat get remembered, and ones that do not get forgotten. Kim describes it simply: “In our system, memory is not updated by a software-defined training algorithm. Instead, it emerges from the timing relationship between physical signals, mimicking the way biological systems learn from repeated experiences. Repeatedly correlated inputs are reinforced, while unrelated signals naturally fade away.”
“Sometimes the simpler the rule, the more powerful it is,” Yang said. “It can be used anywhere.”
Detecting lightning, before it starts a fire
To demonstrate a practical use case, the team ran simulations of a network of sensor fusion units deployed across a wide outdoor area to detect and locate lightning strikes. Lightning is a significant ignition source for wildfires, particularly in remote California terrain where early detection is hardest.
Lightning produces two signals that travel at very different speeds. Light arrives almost instantaneously. Sound takes roughly three seconds to travel each kilometer. When both signals reach the same sensing unit, the gap between them encodes the distance to the strike, which gets stored directly in the device. No server, no wifi signal, no data logger needed. In simulation, the system accurately reconstructed the locations of lightning strikes across a wide area using only what the devices had learned on their own, devices cheap enough and energy-independent enough to sit in the field for years without maintenance.
Yang points out that the same architecture works for wildfire detection more broadly. “You can sense temperature, chemicals, light,” he said. “The system will do the same thing. It will use light and heat to power itself, collect those signals, and learn from them. The philosophy is the same. You just change the sensors, you use it for a different purpose.”
From smart glasses to deep space
The applications extend to anywhere that electronics need to work without infrastructure. Yang sees wearable devices as one natural home for this technology. Smart glasses in particular have struggled for a decade because of weight, battery life, and the need to offload processing to a phone.
“I feel in the future, glasses could be replacing your phone,” Yang said. “You want it to translate Japanese on the street, recognize who you are talking to and process your environment in real time. But you cannot make it heavier than 50 grams, and you cannot have it recharge every hour. The first generation of Google Glass only lasted about 30 minutes with video recording.”
A system like this one could potentially run entirely on the energy harvested from the body or environment. “Everything can be done within whatever you wear,” Yang said. “You harvest pressure, temperature, vibration as the energy source. Ideal case, you do not need a battery. You do not need your phone or a data center to help.” Kim put it simply: “The device does not need to continuously send raw sensory data to a phone, cloud server, or external processor. The sensing, encoding, and learning can happen where the signal is generated. It can process their AI in their own systems, locally. “
The same logic applies to more extreme settings. Space probes, deep-ground sensors, ocean-floor instruments — all settings where power is scarce, communication is slow, and processing needs to happen on site. This also connects to the lab’s earlier work on high-temperature memristors, which operate reliably above 700 degrees Celsius. Combining the two could produce a system that learns autonomously in environments that destroy conventional electronics entirely. “AI can do good things for us in places we cannot go, places that are too hot, too radioactive, too remote,” Yang said.
The most efficient computer already exists
The research was carried out as part of the CONCRETE Center, short for Center of Neuromorphic Computing under Extreme Environments, a Center of Excellence at USC sponsored by the Air Force Office of Scientific Research and the Air Force Research Laboratory, with additional support from the Army Research Office and the National Science Foundation.
As data centers worldwide push against the limits of how much power grids can supply, Yang argues that the answer has been sitting inside every human skull for millions of years.
“Energy is pretty much the key issue for the data center,” he said. “Everything is constrained by energy. And for edge computing, you always want your device to do more with the same amount of energy. You learn from the most efficient computer in the world, which is our brain.”
Published on June 2nd, 2026
Last updated on June 2nd, 2026