Energy-Constrained Information Storage on Memristive Devices in the Presence of Resistive Drift

TL;DR

This work reframes memristor-based information storage under resistive drift as a delay‑conditioned communication problem and proposes a DeepJSCC solution that is explicitly trained with an energy constraint. A differentiable resistive drift channel, modeled via a delay‑conditioned cGAN, enables end‑to‑end optimization, while a novel delay‑conditioned generalised divisive normalisation (cGDN) allows explicit conditioning on recovery delay. The authors derive an energy cost function linked to memristor programming, implement resistance and energy regularisation, and introduce delay specialization to balance energy use across delays. Empirical results on CIFAR‑10 show energy‑aware, delay‑adaptive encodings achieve improved reconstruction across a range of storage delays, and ground‑truth evaluation confirms the differentiable channel model captures drift statistics well, supporting practical applicability for energy‑constrained neuromorphic storage.

Abstract

In this paper, we examine the problem of information storage on memristors affected by resistive drift noise under energy constraints. We introduce a novel, fundamental trade-off between the information lifetime of memristive states and the energy that must be expended to bring the device into a particular state. We then treat the storage problem as one of communication over a noisy, energy-constrained channel, and propose a joint source-channel coding (JSCC) approach to storing images in an analogue fashion. To design an encoding scheme for natural images and to model the memristive channel, we make use of data-driven techniques from the field of deep learning for communications, namely deep joint source-channel coding (DeepJSCC), employing a generative model of resistive drift as a computationally tractable differentiable channel model for end-to-end optimisation. We introduce a modified version of generalised divisive normalisation (GDN), a biologically inspired form of normalisation, that we call conditional GDN (cGDN), allowing for conditioning on continuous channel characteristics, including the initial resistive state and the delay between storage and reading. Our results show that the delay-conditioned network is able to learn an energy-aware coding scheme that achieves a higher and more balanced reconstruction quality across a range of storage delays.

Figures (10)

• Figure 1: In traditional wireless communications information is transmitted through space, whereas, in the case of storage, transmission is through time. We present a DeepJSCC approach to learning to store semantic information on memristive devices, under energy constraints, in the presence of resistive drift noise.
• Figure 2: Costs associated with bringing the resistance of a memristor to a specified value, starting from $R(0) = 500k\Omega$, for $\tau_{\text{final}}=1.0 s$, $K=2.0$, and $R(\tau_{\text{final}})=0.1k\Omega$. In terms of resistance (left), and conductance (right).
• Figure 3: Histogram of the normalised delays following the logarithmic transformation used for training and conditioning the autoencoder.
• Figure 4: The cGDN transform for conditioning the encoder and/or decoder network on the delay, applied to the latent channels outputs for each convolutional layer for the encoder/decoder conditioned settings.
• Figure 5: The DeepJSCC autoencoder architecture for delay-conditioned data storage. To introduce delay conditioning, we replace the GDN layers in bourtsoulatzeDeepJointSourceChannel2019 with cGDN layers, as well as introducing fully connected residual delay processors at the output of the encoder and/or the input of the decoder. cGDN and the delay processors are only included for the encoder/decoder, respectively, if the setting includes delay conditioning. In the diagram, $c_i$ are the number of colour channels in the image, and $c_l$ are the number of latent channels. Changing $c_l$ modifies the compression rate of the autoencoder.