Tunable Synaptic Working Memory with Volatile Memristive Devices

TL;DR

This work presents a demonstration of WM using a silver-based memristive device whose key parameters, retention time and switching probability, can be electrically tuned and adapted to the task at hand, and demonstrates associative symbolic WM.

Abstract

Different real-world cognitive tasks evolve on different relevant timescales. Processing these tasks requires memory mechanisms able to match their specific time constants. In particular, the working memory utilizes mechanisms that span orders of magnitudes of timescales, from milliseconds to seconds or even minutes. This plentitude of timescales is an essential ingredient of working memory tasks like visual or language processing. This degree of flexibility is challenging in analog computing hardware because it requires the integration of several reconfigurable capacitors of different size. Emerging volatile memristive devices present a compact and appealing solution to reproduce reconfigurable temporal dynamics in a neuromorphic network. We present a demonstration of working memory using a silver-based memristive device whose key parameters, retention time and switching probability, can be electrically tuned and adapted to the task at hand. First, we demonstrate the principles of working memory in a small scale hardware to execute an associative memory task. Then, we use the experimental data in two larger scale simulations, the first featuring working memory in a biological environment, the second demonstrating associative symbolic working memory.

Figures (15)

• Figure 1: Conceptual illustration of information storage in working memory. The neural network can store and recall features of an item. The volatile nature of the synapses allows the memory of the stored features to fade in time. After depletion of the memory, the features of a different item can be stored without significant interferences.
• Figure 2: Ag-based volatile memristive device characterization. (a) Sketch of the one-transistor / one-resistor (1T1R) RRAM device together with its working principle. The memristive device (1R) is based on a W / C / 10 nm HfO$\mathrm{_2}$ / Ag stack. The RRAM shows a volatile behavior, i.e., a set operation together with a spontaneous switch off. (b) Time characterization of the retention of the filament. After a 5 V amplitude triangular pulse to switch the cell on with an I$\mathrm{_{CC}}$ = 20 µ A, a constant reading voltage of -150 mV is applied to monitor the retention. (c) Retention time distributions at different compliance currents. The median value of the retention time increases with the compliance current. Inset: applied programming pulse. (d) Switching probability of the device for a single pulse as a function of the amplitude and the pulse duration of the programming pulse. Shorter pulses required higher voltage amplitudes to switch ON. (e) Probability of switching the RRAM to the depending on the number of pulses and their amplitude (1 ms pulses). The circles are the experimental data while the solid lines is the fitting. (f) Effect of the number of pulses and the voltage amplitude (1.6 V) on the switching of the device. The switching of the device is stochastic. Considering a group (burst) of pulses, the probability that the device is in the inside the group increases.
• Figure 3: store and recall experiment. (a) High-level sketch of the working memory. (b) Schematic of the implementation: 5 volatile 1T1R memristive devices are arranged in parallel configuration. The gate is chosen to set I$\mathrm{_{CC}}$ = 17 µ A, that corresponds to a retention time of 28 ms. (c) Color - pattern encoding. (d) Store and recall experiment. During the store phase, a single pattern is fed to the network. Top colored plot: input stimuli. For ease of visualization, each pattern is colored as the color it encodes. Black dots in the bottom part of the upper plot indicate the stored pattern. Bottom plot: measure current fed to the post-neuron. The current threshold for recognition is indicated as a dashed black horizontal line. The traces are cropped on the x-axes to better highlight the salient events. (e) Correlation plot between the expected and measured currents based on the difference between the presented and the stored pattern. Results obtained from 10 different store and recall experiments with P$\mathrm{_{ON}}$ = 5% and stimulation frequency f$\mathrm{_{stim}}$ = 50 Hz. (f) Accuracy of the system in distinguishing the stored pattern under different stimulation and switching conditions. (g) Average current error, defined as the difference between the measured current and the expected current, during 100 patterns applied for the different conditions.
• Figure 4: Large-scale simulation of . (a) Illustration of the network model. 5 different memory items (A,B,C,D and E) can be stored in a recurrent network of spiking neurons. Corresponding strongly connected populations within the network transiently store memory items after activation. (b) Network activity of the model. Black dots show individual spikes of input (top) and network (bottom) neurons. Multiple phases of store and recall are shown. Insets show average firing rates (spikes per second in Hz) over recall phases. Data obtained using a current compliance of 330 µ A, corresponding to average retention times of 1.5 s, and a voltage amplitude of 0.5 V, corresponding to a switching probability of 5%, were used in this simulation. Network behavior using (c) different retention time distribution (change of compliance current) and (d) different switching probability (change of applied voltage amplitude).
• Figure 5: Associative symbolic with volatile memristive devices. (a) Illustration of the network for associative symbolic . (b) Sequence of store and recall. Store/recall input (top row) and neuron output spikes (bottom row) are shown. Inserted pictograms represent the decoded objects and recall queries. Store/recall inputs had a one-to-one fixed connectivity to neurons. After storing an association it can be recalled by cuing the network with an arbitrary memory element. (c) Decoding error plotted as a function of time delay between store and first recall.