Physical neural networks made of analog resistive switching processors are promising platforms for analog computing. State-of-the-art resistive switches rely on either conductive filament formation or phase change. These processes suffer from poor reproducibility or high energy consumption, respectively. Herein, we demonstrate the behavior of an alternative synapse design that relies on a deterministic charge-controlled mechanism, modulated electrochemically in solid-state. The device operates by shuffling the smallest cation, the proton, in a threeterminal configuration. It has a channel of active material, WO 3. A solid proton reservoir layer, PdH x , also serves as the gate terminal. A proton conducting solid electrolyte separates the channel and the reservoir. By protonation/deprotonation, we modulate the electronic conductivity of the channel over seven orders of magnitude, obtaining a continuum of resistance states. Proton intercalation increases the electronic conductivity of WO 3 by increasing both the carrier density and mobility. This switching mechanism offers low energy dissipation, good reversibility, and high symmetry in programming.

In a previous work we have detailed the requirements for obtaining maximal deep learning performance benefit by implementing fully connected deep neural networks (DNN) in the form of arrays of resistive devices. Here we extend the concept of Resistive Processing Unit (RPU) devices to convolutional neural networks (CNNs). We show how to map the convolutional layers to fully connected RPU arrays such that the parallelism of the hardware can be fully utilized in all three cycles of the backpropagation algorithm. We find that the noise and bound limitations imposed by the analog nature of the computations performed on the arrays significantly affect the training accuracy of the CNNs. Noise and bound management techniques are presented that mitigate these problems without introducing any additional complexity in the analog circuits and that can be addressed by the digital circuits. In addition, we discuss digitally programmable update management and device variability reduction techniques that can be used selectively for some of the layers in a CNN. We show that a combination of all those techniques enables a successful application of the RPU concept for training CNNs. The techniques discussed here are more general and can be applied beyond CNN architectures and therefore enables applicability of the RPU approach to a large class of neural network architectures.

No abstract

Nanoscale ionic programmable resistors for analog deep learning are 1000 times smaller than biological cells, but it is not yet clear how much faster they can be relative to neurons and synapses. Scaling analyses of ionic transport and charge-transfer reaction rates point to operation in the nonlinear regime, where extreme electric fields are present within the solid electrolyte and its interfaces. In this work, we generated silicon-compatible nanoscale protonic programmable resistors with highly desirable characteristics under extreme electric fields. This operation regime enabled controlled shuttling and intercalation of protons in nanoseconds at room temperature in an energy-efficient manner. The devices showed symmetric, linear, and reversible modulation characteristics with many conductance states covering a 20× dynamic range. Thus, the space-time-energy performance of the all–solid-state artificial synapses can greatly exceed that of their biological counterparts.