What gets switched
The study published in Nature Communications uses the intrinsic coupling between ferroelectric polarisation and quantum geometry in few-layer WTe2. That coupling lets the second- and third-order nonlinear anomalous Hall effects be switched deterministically, electrically and non-volatilely at room temperature, with the two orders moving in a correlated way.[1]
The load-bearing part is that the two orders switch in opposite directions. Obtaining a complementary pair requires no second material and no second device type; two opposite behaviours sit in the same crystal. The team uses this to build a complementary convolution kernel at the hardware level.[1]
The two numbers
The reported endurance figures are about 10,000 cycles and about 100,000 seconds of retention at room temperature. These are the two quantities that say how long and how many times a device stays reliable, and they are what bound the paper's claim. The third number is the 98 per cent accuracy the convolution kernel reaches on a texture-recognition task.[1]
An accuracy figure from a single task measures the kernel's fit to that task. The gain may come from the complementary kernel itself, and it may equally come from the task's directional structure, since the problem the complementary kernel addresses is the directional specificity of ordinary convolutional networks. The way to separate the two is to repeat the same measurement on a directionally symmetric test.[1]
The steps that remain
The distance between a behaviour shown in one crystal and a manufacturable component is measured in device-to-device variation. If the mechanism survives that variation, an independent group should report the same correlated switching from its own fabrication. That is the signal to watch: a second published measurement of correlated second- and third-order switching in few-layer WTe2 at room temperature.[1]