Source-linked AI summary
Self-Adaptive Spike-Time-Dependent Plasticity of Metal-Oxide Memristors
M. Prezioso, F. Merrikh-Bayat, B. Hoskins, K. Likharev, D. Strukov
TL;DR
Memristor plasticity depends on the device’s initial conductance, raising concerns about external tuning. The paper demonstrates STDP in Al2O3/TiO2-based memristors and models weight dynamics, finding self-adaptation toward a stable conductance distribution.
Problem
Memristor plasticity depends on the initial state, potentially requiring continuous external tuning of each device.
Method
The study implements STDP with multiple spike and window shapes and analyzes conductance changes as functions of initial conductance and spike-time difference.
Results
Self-adaptive STDP drives conductances toward a stable bell-curve distribution independent of their initial values.
Takeaways & Limitations
Al2O3/TiO2-based memristors can implement STDP with window shapes similar to those demonstrated experimentally.
Takeaways & Limitations
The data show slight deviations, including window time asymmetry and a weak, broad second anomaly.
Abstract
from arXiv · showhide
Metal-oxide memristors have emerged as promising candidates for hardware implementation of artificial synapses - the key components of high-performance, analog neuromorphic networks - due to their excellent scaling prospects. Since some advanced cognitive tasks require spiking neuromorphic networks, which explicitly model individual neural pulses (spikes) in biological neural systems, it is crucial for memristive synapses to support the spike-time-dependent plasticity (STDP), which is believed to be the primary mechanism of Hebbian adaptation. A major challenge for the STDP implementation is that, in contrast to some simplistic models of the plasticity, the elementary change of a synaptic weight in an artificial hardware synapse depends not only on the pre-synaptic and post-synaptic signals, but also on the initial weight (memristor's conductance) value. Here we experimentally demonstrate, for the first time, STDP protocols that ensure self-adaptation of the average memristor conductance, making the plasticity stable, i.e. insensitive to the initial state of the devices. The experiments have been carried out with 200-nm Al2O3/TiO2-x memristors integrated into 12x12 crossbars. The experimentally observed self-adaptive STDP behavior has been complemented with numerical modeling of weight dynamics in a simple system with a leaky-integrate-and-fire neuron with a random spike-train input, using a compact model of memristor plasticity, fitted for quantitatively correct description of our memristors.
Results
Experiments show that metal-oxide memristors implement biologically similar STDP windows whose conductance changes depend strongly on initial state. A fitted compact model predicts self-adaptation toward stable mid-range conductances in a simple spiking network.
- Experimental setup: The experiments used Pt/Al2O3/TiO2-x/Ti/Pt memristors integrated into 12×12 crossbars and applied three biologically plausible spike-waveform protocols.Each conductance-change measurement was repeated 10 times after resetting the device to the same initial state.
- Experimental results: The three spike shapes produced three representative STDP window shapes found in various biological synapses.Changing switching polarity and/or reversing the sign of Δt can also produce anti-Hebbian window behavior.
- Initial-state dependence: Across the memristors’ dynamic range, conductance increases vanish near maximum conductance and decreases vanish near minimum conductance because switching saturates.This demonstrates strong dependence of plasticity on the initial conductance, which could otherwise require continuous external tuning of each device.
- Modeling: A compact phenomenological model with fitted parameters describes the experimentally observed conductance changes reasonably well.The model captures dependence on both initial conductance and spike timing, while its global response is not self-similar across Δt.
- Implications and caveats: The modeled self-adaptation supports stable operation, whereas additive STDP models with ΔG independent of G0 cannot ensure such self-adaptation.The authors note a weak, broad second peak in the simulated final-conductance distribution and slight waveform deviations, including time asymmetry in one window.
Supplementary Information
The supplementary information details fabrication and electrical characterization of 200-nm crossbar memristors, then describes simulation procedures for conductance-distribution dynamics.
- Device fabrication: The crossbar lines are 200 nm wide with 400 nm gaps, and the completed dies are annealed before wire bonding.Annealing uses 10% H2 and 90% N2 for 15 minutes at 300°C.
- Device fabrication: The devices use patterned Ta/Pt bottom wires, an Al2O3 barrier, and a TiO2 switching layer in a crossbar structure.The film stack includes 5-nm Al2O3 and 30-nm TiO2; the final crossbar layout is shown in Fig. 1a.
- Electrical characterization: The measured switching thresholds are close to 0.7 V for set and -0.8 V for reset operations.Set switches from low to high conductive state, while reset is performed at the negative threshold.
- Simulation analysis: Supplementary simulations vary the fitting constant a- in the STDP model and use a vertical slice of the conductance-distribution map for a- = 74.52.Other fitting constants and model assumptions match those specified in the main text and Fig. 3b.