Building time-surfaces by exploiting the complex volatility of an ECRAM memristor

TL;DR

The paper investigates how volatile, three-terminal Li_xWO3 memristors with tunable short-term plasticity and dual exponential decay can be used for temporal computation in neuromorphic hardware. It derives a stochastic and an ideal model for the memristor dynamics, integrates them into the Hierarchy of Time Surfaces (HOTS) architecture, and evaluates recognition performance on event-based datasets NMNIST and POKERDVS. Key findings show that device stochasticity has minimal impact on accuracy, while STP and dual time constants improve performance by enabling multi-scale temporal integration; programmable pulse settings allow tuning to dataset-specific temporal statistics, achieving competitive results with limited training. The work demonstrates a practical route to harness memristor temporal dynamics for compact, energy-efficient neuromorphic systems and informs design choices for matching device dynamics to problem timescales.

Abstract

Memristors have emerged as a promising technology for efficient neuromorphic architectures owing to their ability to act as programmable synapses, combining processing and memory into a single device. Although they are most commonly used for static encoding of synaptic weights, recent work has begun to investigate the use of their dynamical properties, such as Short Term Plasticity (STP), to integrate events over time in event-based architectures. However, we are still far from completely understanding the range of possible behaviors and how they might be exploited in neuromorphic computation. This work focuses on a newly developed Li$_\textbf{x}$WO$_\textbf{3}$-based three-terminal memristor that exhibits tunable STP and a conductance response modeled by a double exponential decay. We derive a stochastic model of the device from experimental data and investigate how device stochasticity, STP, and the double exponential decay affect accuracy in a hierarchy of time-surfaces (HOTS) architecture. We found that the device's stochasticity does not affect accuracy, that STP can reduce the effect of salt and pepper noise in signals from event-based sensors, and that the double exponential decay improves accuracy by integrating temporal information over multiple time scales. Our approach can be generalized to study other memristive devices to build a better understanding of how control over temporal dynamics can enable neuromorphic engineers to fine-tune devices and architectures to fit their problems at hand.

Figures (7)

• Figure 1: The structure and operation of the physical Li$_\textbf{x}$WO$_\textbf{3}$ electrochemical synapse modelled in this paper. (a) The Li$_\textbf{x}$WO$_\textbf{3}$ electrochemical synapse is a three-terminal device with a (S)ource, (G)ate and (D)rain. The gate and channel between S and D are built by deposition of tungsten oxide (WO$_\textbf{3}$) films (red) on a LaALO$_\textbf{3}$ substrate (green). The films are connected to gold terminals (yellow). Lithium ($\mathrm{Li+}$) and $\mathrm{ClO_4-}$ ions are introduced by applying a drop of electrolyte gel (blue) on top of the device. (b) The electrical behavior of the memristor in response to a square WRITE voltage pulse applied between (G) and (D) and a small DC READ voltage (0.1 V) applied between (S) and (D). (c) The electrochemical behavior of the memristor.
• Figure 2: Example response of our model to a pulse at $t_i$ with pulse width $w$ assuming initial steady state (in green). The model presents a linear increase (in blue) for the duration of the pulse. At its end, it will relax to a steady state following a double exponential decay (in gold).
• Figure 3: Our stochastic model (in orange) and a memristor recording (in blue) for a pulse of 1V 200us. We sample the response parameters from the parameter distributions obtained from fitting our model to experimental data.
• Figure 4: Top row (a-f): Mapping spatio-temporal event patterns into Time-Surface features using memristive synapses. Bottom row (g-m): N-MNIST classification using the memristive HOTS network. An input digit (a) is presented to the event-based sensor (b), which produces asynchronous event streams (c) at each pixel based on the luminance variation over time. Each pixel stream is input to a memristor (d), which interpolates the spike train with exponential decay kernels (in red) with time constants $\tau_1$ and $\tau_2$, resulting in time-varying conductances (e). For each new event $ev_i$ (indicated in light blue), we generate a Time-Surface by sampling the conductances from memristors at neighboring pixels, resulting in a 2D map that encodes temporal correlations between events at different pixels (f). The spiking activity (g) from the event-based sensor has two polarities, indicating increases (orange) or decreases (blue) in luminance. This figure shows the polarity of the last spike at each pixel, if any, during the last 10ms before a reference event $ev^1_i$. Time surfaces (h) are created by sampling changes to memristor conductances in a square neighborhood around the reference event, shown in red. Time surfaces are clustered (i). Input events produce output events (j) with the same location and time stamp, but polarity determined by the closest cluster index. This new set of events is spatially sub-sampled, resulting in a larger effective neighborhood size. In the next layer, these are used as inputs to generate new time surfaces (k) that are again mapped to clusters (l) to produce output events (m). This process can be repeated multiple times to increase the temporal and spatial scale of the classified features.
• Figure 5: (a) Recognition rates of the Ideal and Noisy networks with the Support Vector and Euclidean Classifiers. (b) A qualitative comparison between Time-Surfaces computed over the entire input using the Ideal and Noisy memristor models. (c and d) The effect of Time-Surface perturbation on cluster (i.e., polarity) assignment in Layers 1 and 2. Blue dots indicate events where the Ideal and Noisy networks make the same cluster assignment. Otherwise, the Ideal (orange dots) and Noisy (green dots) assign events to different clusters. (e) We call this effect cluster dislocation. (f and g) The Mutual Information between the cluster response and the N-MNIST digit labels for Layers 1 and 2 at different Temporal Integration scales. (h) The Mutual Information Percentage Loss due to cluster dislocation. The effect of dislocation is small and constant across all timescales, except for a singularity when the temporal window is zero due to division by zero MI when computing the percentage. More importantly, cluster dislocation is equally likely to increase or decrease Mutual Information.