�鶹��Ժ

Spin-based memory advance brings brain-like computing closer to reality

Researchers at National Taiwan University have developed a new type of spintronic device that mimics how synapses work in the brain—offering a path to more energy-efficient and accurate artificial intelligence systems.

In a study in Advanced Science, the team introduced three novel memory device designs, all controlled purely by electric current and without any need for an external magnetic field.

Among the devices, the one based on "tilted anisotropy" stood out. This optimized structure was able to achieve 11 stable memory states with highly consistent switching behavior.

Unlike conventional memory devices that often rely on unstable physical processes, this device demonstrated excellent reliability with a cycle-to-cycle variation of just 2%. This makes it ideal for simulating how synapses in the brain strengthen or weaken connections over time—critical for machine learning.

The team went a step further by simulating how this multi-state memory could be used in real neural networks. They applied the device's memory states to a convolutional neural network model (ResNet-18) for image classification. Using a process called post-training quantization, they matched the physical memory states to digital weights in the AI model.

The results were impressive: with the 11-state device, classification accuracy reached up to 81.51%—nearly the same as the original uncompressed software model.

These findings show that spintronic synapses could power future neuromorphic chips that are not only faster and more compact but also far more energy-efficient.

"Our work paves the way for reliable, scalable neuromorphic computing using purely electrical control," said Prof. Chi-Feng Pai.