講演情報
[8a-N304-5]Superparamagnetic Tunnel Junction-based Probabilistic Neuron with Colored Noise
〇Zhiqiang Liao1, Deyi Wang1, Hitoshi Tabata1 (1.Univ. Tokyo)
キーワード：
Superparamagnetic tunnel junction、Colored noise、Probabilistic computing
Superparamagnetic tunnel junctions (SMTJs) are nanoscale devices deliberately operated in the superparamagnetic regime, where the energy barrier is so low that thermal fluctuations drive spontaneous switching between parallel and antiparallel magnetization states. At room temperature they exhibit stochastic telegraph switching that follows Neel-Arrhenius statistics, with rates and state probabilities finely tunable by tiny voltages through spin-transfer or spin-orbit torques, which makes them effectively act as controllable probabilistic bits. This intrinsic randomness, coupled with nanoscale size and CMOS compatibility, makes SMTJs compelling primitives for sub-threshold neuromorphic computing that promise ultra-low energy consumption.
Probabilistic neurons based on SMTJs are typically analyzed under the assumption of a white-noise environment. In practice, however, real-world noise exhibits nonuniform power spectral densities, which is analogous to the spectra of colored light. To address this gap, we theoretically investigate SMTJ dynamics and the resulting probabilistic activation functions under various colored noises. For equal overall noise intensity, low-frequency-weighted red and pink noise produce probability outputs with gentler activation gradients, whereas high-frequency-weighted blue and violet noise yield markedly steeper gradients. Moreover, the activation functions obtained under red and pink noise display larger statistical uncertainties than those under other noise types. These results imply that low-frequency-dominated noise can be treated as effectively equivalent to a high-frequency-dominated noise with higher intensity. Our simulation on MNIST and MedMNIST datasets shows that, compared to high-frequency-dominated noise, low-frequency-dominated noise enables SMTJ-based convolutional neural networks to attain peak performance under substantially smaller driving voltage, thereby reducing energy consumption.
Probabilistic neurons based on SMTJs are typically analyzed under the assumption of a white-noise environment. In practice, however, real-world noise exhibits nonuniform power spectral densities, which is analogous to the spectra of colored light. To address this gap, we theoretically investigate SMTJ dynamics and the resulting probabilistic activation functions under various colored noises. For equal overall noise intensity, low-frequency-weighted red and pink noise produce probability outputs with gentler activation gradients, whereas high-frequency-weighted blue and violet noise yield markedly steeper gradients. Moreover, the activation functions obtained under red and pink noise display larger statistical uncertainties than those under other noise types. These results imply that low-frequency-dominated noise can be treated as effectively equivalent to a high-frequency-dominated noise with higher intensity. Our simulation on MNIST and MedMNIST datasets shows that, compared to high-frequency-dominated noise, low-frequency-dominated noise enables SMTJ-based convolutional neural networks to attain peak performance under substantially smaller driving voltage, thereby reducing energy consumption.