Researchers at the National University of Singapore have built a transistor that can filter image pixels before a spiking neural network processes them. The same device can perform logic, act like a neuron or store a synapse-like connection, depending on its electrical signals. In the paper’s system-level model, this approach reduced the hardware resources assigned to image tasks; its energy estimates are not measurements from a complete chip.
Image-processing hardware often sends an entire camera frame into computation, even when a task needs only one region. A spiking neural network (SNN), an AI model whose units communicate through brief pulses, can avoid work when its neurons remain inactive. But unrelated pixels can still trigger activity. In the NUS design, logic gates sit before the SNN and block pixels outside a task-selected region, so only the retained pixels enter the pulse-based network.
The device uses an atom-thin molybdenum disulfide (MoS2) semiconductor channel and a ferroelectric hafnium-zirconium-oxide gate layer. Electrical bias determines its role: steady conditions enable logic; short voltage pulses produce a temporary current response like a leaky neuron; longer pulses trap charge and change conductance, representing a synaptic weight. In the system-level design, the devices are organized into 16-by-16 tiles: blocks of transistors assigned to filtering or to the network’s neuron and synapse functions.
In a device-calibrated simulation of handwritten-digit recognition, cropping MNIST inputs from 28 × 28 to 20 × 20 pixels cut the number of active 16-by-16 tiles by about 36% and the model’s synaptic weights by 48%. Accuracy fell by about 0.3 percentage points from 95.9%; cropping further to 16 × 16 cut into the digits and lowered accuracy to 94.4%. The comparison shows the trade-off: filtering saves resources only while the removed area is not needed by the task.
For a separate modeled traffic scene with a 3,072-pixel input, estimated array-level inference energy fell from 0.29 joules on the full image to 0.12 joules for vehicle recognition and 0.066 joules for license-plate recognition. The estimate multiplies measured per-tile energy costs by simulated tile counts and excludes supporting circuits.
Control signals specify which region each task keeps; the paper does not demonstrate a camera system finding that region automatically. Its result is a hardware-level way to apply a supplied selection before neural computation, rather than a full vision pipeline.
