Neuromorphic Hardware Built a Two-Way Feedback Loop for Temperature Control

A hardware system inspired by an autapse, a neuron’s connection back to itself, has demonstrated localized positive and negative feedback without routing every adjustment through a central processor. The laboratory prototype coupled a solid-electrolyte synaptic transistor to a bidirectional leaky integrate-and-fire neuron, then used temperature control as a concrete test.

Why the feedback loop matters

Most electronic control systems sense a condition, send data to a processor, calculate a response and issue a new command. That architecture is flexible, but it can add data movement, latency and energy use. Biological nervous systems often place feedback closer to the point where signals are integrated. An autapse is an unusually direct example: a neuron’s output reconnects to its own input and changes how it fires next.

The researchers reproduced that principle with two hardware elements. A synaptic transistor accepted several inputs, while an artificial neuron accumulated the resulting signal and fired after crossing a threshold. Feeding the neuron’s output into an extra gate on the synaptic transistor created a closed loop. Depending on how the feedback was configured, it reinforced later firing or suppressed it.

The important result is not that the circuit “thinks” like a brain, but that sensing, integration and corrective feedback can be compressed into a small local hardware loop.

What the experiment showed

In the thermoregulation demonstration, temperature-dependent signals modulated the artificial neuron’s firing rate in both directions. The setup could therefore represent a controller that reacts differently when a system is too hot or too cold. The paper describes this as bidirectional feedback because the same architecture can strengthen or weaken the response rather than applying only one kind of correction.

The solid-electrolyte transistor also multiplexed several inputs, reducing the need for a separate device for every signal. This is relevant to compact edge systems, where sensors and controllers must operate with limited power and where repeatedly communicating with a remote processor can be expensive.

The approach complements other attempts to combine sensing, memory and computation in individual devices. NewTqnia previously covered a vertical transistor that processed motion and stored classifier weights. The new study targets a different function: a self-regulating feedback loop rather than visual classification.

What comes next

The useful next comparison would place this local feedback architecture against a conventional microcontroller on the same task, measuring total energy, response time, accuracy and drift over extended operation. More complex tests would also need multiple coupled loops, because practical machines regulate many variables at once.

If those tests hold up, autapse-like circuits could become building blocks for always-on sensors, adaptive thermal management and compact machines that correct simple conditions locally. For now, the paper establishes the device concept and its bidirectional behavior, not a finished route to broadly autonomous hardware.