BBC Russian

Neuroscience-inspired information-integration system based on stochastic magnetic tunnel junctions

Li Zhao, Meiting Zhang, Yajun Zhang, Yuanyuan Mi, Zhe Yuan, and Ke Xia

Phys. Rev. Applied 21, 064040 – Published 17 June 2024

Abstract

Effective integration of multiple sensory information in the brain is essential to achieve accurate information perception and processing despite the noise and imperfections in biological neural systems, providing strong inspiration for improving computational accuracy in neuromorphic computing. In the future development of brainlike chips, e.g., biomimetic robots, multisensory-information integration is an essential capability for recognizing the external circumstances. Here, based on the computational model proposed by neuroscientists, we propose a hardware implementation scheme of the decentralized multisensory information-integration system using spintronic devices, in which magnetic tunnel junctions are employed as artificial neurons with stochastic dynamics. Using a one-dimensional continuous variable (orientation, head direction, etc.) as a typical example, we demonstrate that the input information from noisy cues is extracted with high accuracy after integration. This spintronic neuromorphic system exhibits remarkable tolerance for both nonuniform devices and the malfunction of different modules. The computational advantages and robustness of the spintronics-based information-integration system provide an important basis for the development of hardware neuromorphic platforms.

Received 26 August 2023
Revised 26 March 2024
Accepted 17 May 2024

DOI:https://doi.org/10.1103/PhysRevApplied.21.064040

Physics Subject Headings (PhySH)

Neuromorphic computing Probabilistic computing Spintronics Stochastic processes

Artificial neural networks Magnetic tunnel junctions

Micromagnetic modeling Neural network simulations

Condensed Matter, Materials & Applied PhysicsInterdisciplinary Physics

Authors & Affiliations

Li Zhao^1,2,3,‡, Meiting Zhang ^4,‡, Yajun Zhang^1,2,3, Yuanyuan Mi ^5,*, Zhe Yuan ^2,3,†, and Ke Xia⁶

¹Center for Advanced Quantum Studies and Department of Physics, Beijing Normal University, Beijing 100875, China
²Institute for Nanoelectric Devices and Quantum Computing, Fudan University, Shanghai 200433, China
³Interdisciplinary Center for Theoretical Physics and Information Sciences (ICTPIS), Fudan University, Shanghai 200433, China
⁴Center for Neurointelligence, School of Medicine, Chongqing University, Chongqing 400044, China
⁵Department of Psychological and Cognitive Sciences, Tsinghua University, Beijing, 100084, China
⁶School of Physics, Southeast University, Nanjing 211189, China

^*Corresponding author: [email protected]
^†Corresponding author: [email protected]
^‡L. Zhao and M. Zhang contributed equally to this work.

Article Text (Subscription Required)

Click to Expand

Supplemental Material (Subscription Required)

Click to Expand

References (Subscription Required)

Click to Expand

Issue

Vol. 21, Iss. 6 — June 2024

Subject Areas

Reuse & Permissions

Access Options

Author publication services for translation and copyediting assistance advertisement

Images

Figure 1
(a) Schematic illustration of population coding. Lines in the left panel indicate the neurons’ tuning curves, which are bell shaped and centered at their preferences. One-dimensional variable $θ_{ext}$ induces activity in a group of neurons, resulting in a bumplike population activity (right panel). Stimulus value $θ_{est}$ is then decoded based on the neural population activity [11]. (b) Switching frequency of the MTJs with different preferences as functions of the input voltage. Frequency is calculated by sampling 100 states within 1 ms. Dashed lines are fitted using a Gaussian function. (c) Schematic of the proposed architecture of a MTJ-based CANN. Every MTJ in the CANN is subjected to a unique local bias to represent the neuron with its own preferred feature, as indicated by the substrate at the bottom. Green arrows denote excitatory connections for the MTJ in the dashed frame, and the connections have translationally invariant symmetry for all the MTJs. Solid magenta line schematically illustrates the bell-shaped response bump of the network under an external stimulus. Inhibitory connections are not explicitly shown. (d) Population activities of the network with and without connections. Solid white lines represent the estimate decoded from the network response. Stimulus is input to the network at $t < 20 Δ t$ . (e) Deviation of the network estimate from the real stimulus and the variance of the network estimate as functions of the system’s noise strength.
Reuse & Permissions
Figure 2
(a) Structure of two coupled CANNs based on MTJs. Two CANNs have the same structure as in Fig. 1. Each CANN receives an independent voltage as the input sensory signal and communicates through the excitatory connections between networks denoted by purple arrows. (b) Schematic diagram of the three external stimulus input conditions: (1) only $V_{ext}^{1}$ is applied to network 1 (left); (2) only $V_{ext}^{2}$ is applied to network 2 (center); (3) $V_{ext}^{1}$ and $V_{ext}^{2}$ are simultaneously applied to the two networks (right). (c) Population activities of two networks for the three stimulus conditions in a temporal order. (d) Decoded $θ_{est}$ from the firing rate in network 1 for 400 trials, while the right curves illustrate the distribution of the 400 estimated values. (e) Average distribution of extracted $θ_{est}$ . (f) Mean variance obtained over 20 runs. Error bars indicate the standard deviation of the variances over 20 runs. Bayesian inference is plotted for comparison. Three asterisks indicate the significant difference (***, p < 0.001).
Reuse & Permissions
Figure 3
Information integration by three networks that are reciprocally connected with one another. Top inset illustrates the system consisting of three CANNs. From left to right, normal; external stimulus for network 3 not received (B1); input communication to network 3 blocked (B2); output communication from networks blocked (B3). Bottom histograms show the calculated average variance from network 1 under the four input conditions: networks 1–3 receive the external stimuli, respectively, and all three networks receive the input simultaneously. Error bars represent the standard deviation over 20 runs, while every run contains 400 independent trials. Variance of Bayesian inference is plotted for comparison. Note that with B1 or B3 and only network 3 receiving the external stimulus, the system does not respond. Three asterisks indicate the significant difference (***, p < 0.001).
Reuse & Permissions

Physical Review Applied

Neuroscience-inspired information-integration system based on stochastic magnetic tunnel junctions

Li Zhao, Meiting Zhang, Yajun Zhang, Yuanyuan Mi, Zhe Yuan, and Ke Xia

Phys. Rev. Applied 21, 064040 – Published 17 June 2024

Abstract

Physics Subject Headings (PhySH)

Authors & Affiliations

Article Text (Subscription Required)

Supplemental Material (Subscription Required)

References (Subscription Required)

Issue

Subject Areas

Access Options

Authorization Required

Other Options

Download & Share

Images

Figure 1

Figure 2

Figure 3

Log In

Search

Article Lookup

Paste a citation or DOI

Enter a citation