Source-linked AI summary
Magnetic Tunnel Junction Mimics Stochastic Cortical Spiking Neurons
Abhronil Sengupta, Priyadarshini Panda, Parami Wijesinghe, Yusung Kim, Kaushik Roy
TL;DR
Hardware spiking-neuron implementations have mainly modeled deterministic behavior, whereas cortical pyramidal neurons spike probabilistically. This paper maps stochastic cortical spiking to thermally induced MTJ switching and demonstrates pattern-recognition functionality with lateral inhibition and homeostasis, alongside low reported energy consumption.
Problem
Existing hardware spiking-neuron implementations mainly focus on deterministic models that do not capture the probabilistic firing and noise-prone computation of biological neurons.
Method
The paper maps probabilistic cortical spiking to thermally noisy stochastic switching in Magnetic Tunnel Junctions and uses MTJ neurons in a network with lateral inhibition and homeostasis.
Results
The MTJ-neuron network learns representative digit models and supports pattern recognition, while the neuron write energy is approximately 1 fJ per neuron per time-step.
Takeaways & Limitations
Stochastic MTJ neurons provide a hardware mapping of cortical probabilistic spiking and can potentially support Bayesian computation similar to the brain.
Takeaways & Limitations
Interfacing memristive synaptic crossbar arrays with memristive neurons may be difficult because of high threshold voltages and high resistance values.
Abstract
from arXiv · showhide
Brain-inspired computing architectures attempt to mimic the computations performed in the neurons and the synapses in the human brain in order to achieve its efficiency in learning and cognitive tasks. In this work, we demonstrate the mapping of the probabilistic spiking nature of pyramidal neurons in the cortex to the stochastic switching behavior of a Magnetic Tunnel Junction in presence of thermal noise. We present results to illustrate the efficiency of neuromorphic systems based on such probabilistic neurons for pattern recognition tasks in presence of lateral inhibition and homeostasis. Such stochastic MTJ neurons can also potentially provide a direct mapping to the probabilistic computing elements in Belief Networks for performing regenerative tasks.
Introduction
Neuromorphic computing seeks hardware implementations that emulate brain mechanisms, while spiking neurons provide a highly bio-faithful model of neural information processing. This paper focuses on mapping probabilistic cortical neuron behavior to hardware.
- Neuromorphic computing emulates underlying neuroscience mechanisms through equivalent hardware implementations.
- Spiking neurons integrate incoming spikes through membrane-potential dynamics and emit output spikes when a threshold is crossed.
- The paper demonstrates a nano-magnetic device that mimics stochastic spiking neurons observed in cortical pyramidal circuits.
Magnetic Tunnel Junction as a spiking neuron
The paper maps biological spiking-neuron dynamics onto a thermally noisy MTJ whose magnetization integrates and leaks input, then switches probabilistically between stable resistance states. A three-terminal implementation separates writing, reading, and resetting, enabling stochastic-neuron networks for pattern recognition.
- Device structure: An MTJ uses a pinned reference layer and a current-controlled free layer separated by a tunneling oxide barrier, with stable parallel and anti-parallel resistance states.The barrier height EB stabilizes both states.
- Neuron dynamics: The free-layer magnetization acts like a membrane potential: it integrates input pulses, leaks without input, and fires when it switches to the opposite stable state.The MTJ firing state is the anti-parallel configuration when initialized in the parallel state.
- Network application: A stochastic network uses resistive crossbar synapses, lateral inhibition, and homeostasis-related learning mechanisms for digit recognition.The network receives rate-encoded image inputs and uses nine excitatory neurons whose learned responses become selective to input classes.
- Stochastic response: Thermal noise makes switching stochastic, and switching probability increases with input-current magnitude while remaining dependent on barrier height and pulse duration.The probability curve shifts with barrier height and becomes more dispersed as pulse width decreases.
- Neuron circuit: A three-terminal MTJ neuron decouples synaptic write current from read current, then resets the device after a detected spike.The reset delay produces a period during which the neuron cannot generate another spike, analogous to a refractory period.
Spiking Neural Network based on MTJ neurons
The network uses stochastic MTJ neurons with lateral inhibition, homeostasis, programmable synapses, and STDP to learn digit patterns. Learning produces digit-selective spiking responses while exploiting changing switching probabilities and low per-step energy consumption.
- Network input: 100 images of digits “0” and “1” were presented as Poisson spike trains whose frequencies were proportional to pixel intensity.The network was simulated for multiple time steps per image.
- Learning mechanisms: Lateral inhibition suppresses non-spiking neurons temporarily, promoting selective responses to input patterns during learning.A common inhibitory signal acts for a period τinh after a neuron spikes.
- Learning mechanisms: Homeostasis scales MTJ input current as learning progresses to prevent individual neurons from dominating spiking activity.The resulting switching probability decreases and neural and learning events become sparser.
- Network architecture: Programmable resistive crossbar synapses used 4-bit discretization with a maximum-to-minimum resistance ratio of 20.The architecture can use phase-change devices, Ag-Si memristors, or spintronic synapses for synaptic functionality.
- Learning outcome: After a few epochs of unsupervised STDP learning, neurons learned representative digit models and became trained to respond to specific digits.Spiking patterns were used to detect the input image class.
- Energy consumption: The average MTJ write energy was ~1 fJ per neuron per time-step, while read and reset energies were ~1.6 fJ and ~4.5 fJ, respectively.The write and reset cycles were each 0.5 ns long; reset energy was used only when the neuron required resetting.
Conclusions
The paper positions stochastic MTJ switching as a route to probabilistic spiking neurons, emphasizing interface advantages and potential neuromorphic applications. It also identifies practical constraints in interfacing memristive synaptic arrays with memristive neurons.
- Spiking neurons provide a more biologically realistic basis for low-power, high-accuracy recognition platforms than threshold-only neural models.
- The work demonstrates mapping MTJ stochastic switching under thermal noise to probabilistic spiking neurons, extending prior stochastic-device use beyond synaptic learning.
- Low terminal voltage and decoupled read/write current paths support interfacing MTJ neurons with resistive synaptic crossbar arrays.
- Memristive synaptic crossbars may be difficult to interface with memristive neurons because memristive devices typically have threshold voltages > 1V and resistances of KΩ-MΩ.
- Noisy MTJ switching can potentially support probabilistic neural platforms for Bayesian computation and architectures such as Restricted Boltzmann Machines and Deep Belief Networks.
Author contributions statement
The research was conceived by Sengupta and Roy, with simulation, device-framework, and writing contributions distributed across the author team.
- Sengupta conceived the study, wrote the paper, developed the simulation framework, and performed the simulations.
- Panda and Wijesinghe contributed equally and assisted with the simulations.
- Kim helped develop the device simulation framework, while Roy supported writing, concept development, and discussion of results.