Hardware-Aware On-Chip Learning with a Scalable Single-Transistor-based Neuromorphic Platform

Modern computers are incredibly powerful, but they are also hungry. As artificial intelligence grows more complex, the energy required to run it has become a bottleneck, especially for devices that need to run for long periods on small batteries, like robots or sensors in the wild. To solve this, scientists have turned to a different way of thinking about computing, one that mimics the human brain. Instead of processing information in a steady, continuous stream like a traditional calculator, these "neuromorphic" systems use tiny, asynchronous bursts of information called spikes. This method is far more efficient because the system only works when there is something new to process. However, teaching these brain-like machines to learn new tasks while they are running has been a major hurdle. In most designs, the internal state of a neuron—its "memory" of what it has just received—is hidden inside the hardware, making it nearly impossible for the system to adjust its own connections and learn from mistakes without expensive, energy-draining external sensors.

A team of researchers at Seoul National University and the Korea Institute of Science and Technology has now demonstrated a way to teach these systems without ever needing to peek inside the neurons. They have built a new type of chip where the learning process relies only on signals that are already visible on the surface: the electrical current flowing between components and the timing of the spikes themselves. By designing the hardware and the learning software to work together, they created a system that can learn on the chip, using a single transistor to act as a synapse, a neuron, and a learning signal generator all at once. This approach removes the need to read the hidden internal voltage of a neuron, a step that previously required bulky, power-hungry circuitry. The result is a unified platform that is not only smaller and simpler but also capable of learning continuously, adapting to new information while remembering what it has already learned.

The core of this breakthrough lies in how the researchers solved the problem of "state access." In a traditional artificial brain, to teach a neuron to fire at the right time, the computer needs to know exactly how much electrical charge is sitting inside it at any given moment. In physical hardware, this charge is trapped inside the device, and measuring it requires complex, energy-intensive circuits that defeat the purpose of having a low-power system. The researchers realized they did not need to measure this hidden charge if they could use the current flowing into the neuron as a proxy. They developed a method called a current-based surrogate gradient. Instead of asking "how full is the neuron's battery?", the system asks "how much current is flowing right now?" If the current is within a specific, healthy range, the system knows the neuron is learning correctly. If the current is too high or too low, it knows an adjustment is needed. This allows the chip to generate its own learning signals using only the data that is naturally available at the connection points between neurons.

To make this work in the real world, the team fabricated a special chip using a standard manufacturing process, but with a clever twist. They used a single type of transistor that could perform three different jobs depending on how it was configured. One group of these transistors acted as synapses, the connections that store the strength of a memory. Another group acted as neurons, which accumulate electrical charge until they reach a tipping point and fire a spike. The third group acted as the learning mechanism, watching the current and deciding when to send a correction signal. Because all three components are built from the same basic building block, they can be packed together densely on a single chip. The researchers showed that these transistors could reliably store different levels of weight, integrate incoming signals over time, and fire spikes in a pattern that mimics biological neurons. Crucially, the learning component could detect when the system was operating correctly or incorrectly just by looking at the current, without needing to access the hidden internal voltage of the neuron.

The team tested this system through extensive computer simulations that used the actual measured behavior of their physical chips. They trained the network on a variety of tasks, ranging from recognizing handwritten digits to identifying complex objects in images and interpreting video streams from event-based cameras. In these tests, the new method performed just as well as traditional learning methods that rely on full access to internal states. The system was able to learn new classes of objects without forgetting the ones it had already mastered, a capability known as continual learning. This is a significant achievement because many previous attempts to teach these systems on the chip resulted in "catastrophic forgetting," where learning something new would erase old memories. By using their new learning signals, the researchers showed that the system could adapt to new information while retaining its previous knowledge, even when the hardware itself was imperfect and noisy.

The implications for energy efficiency are substantial. The researchers estimated that by using their analog, spike-based approach, the energy required to run an inference—making a decision based on new data—could be reduced by more than 60 percent compared to fully digital systems. In some specific configurations, the savings were even higher, reaching over 85 percent. This is because the system avoids the massive energy cost of converting analog signals to digital numbers and back again, and it only consumes power when a spike actually occurs. While the process of teaching the chip initially still requires some high-voltage pulses that consume energy, the researchers noted that this cost is a one-time setup or a rare event, whereas the energy savings during daily operation are continuous and significant.

This work does not claim to have solved every problem in artificial intelligence, nor does it suggest that these chips are ready to replace the processors in your smartphone tomorrow. The researchers were careful to point out that while the learning signals work well, the full energy cost of a complete, integrated learning system has not yet been measured in a real-world device. There are still challenges in scaling up the manufacturing and integrating the control circuits needed to manage these chips. However, the study provides a clear proof of concept that it is possible to build a brain-like computer that learns on its own without needing to see its own internal thoughts. By designing the hardware to speak the same language as the learning algorithm, the team has opened a path toward adaptive, energy-efficient machines that can learn and evolve in the real world, just like the biological brains they are designed to emulate.